AIニュース 2026-06-30

Recursive Self-Evolving Agents via Held-Out Selection

LLM agents are increasingly improved without weight updates by evolving a natural-language artifact, such as reflections, workflows, playbooks, cheatsheets, or optimized prompts, that conditions a frozen policy. Such methods are typically reported as wins on the single benchmark where they help. We study them apples-to-apples and surface a sharper picture. We introduce RSEA, a Recursive Self-Evolving Agent that carries a compact three-layer natural-language state: an imperative strategy, reusable skills, and a procedural playbook. Across generations, RSEA rewrites all three layers from its own trajectories and commits a candidate only if it does not regress on a disjoint held-out split, using a strict keep-better gate. Across four diverse benchmarks, ALFWorld, GAIA, (\tau)-bench, and WebShop, and six faithful baselines, ReAct, Reflexion, GEPA, AWM, ACE, and Dynamic Cheatsheet, all evaluated on one shared local backbone, we find three main results. First, no artifact universally wins. RSEA is the strongest single-pass method on ALFWorld, reaching 69.3% compared with 64.6% for ReAct (McNemar (p=0.015)), and reaches 79.4% with retry, the best overall result. However, concrete-workflow induction, represented by AWM, is best on the strong-backbone tool-use tasks. Second, unguarded context evolution is high-variance and unsafe. Dynamic Cheatsheet, which curates context online without a held-out gate, is near-best on ALFWorld at 70.7%, yet collapses on WebShop, with a score of 0.14 compared with 0.43 for ReAct. Third, RSEA's strict held-out selection is what makes recursive self-evolution monotone-safe: it never significantly underperforms the base agent on any benchmark and falls back to vanilla ReAct when evolved context would hurt.

Data and Evaluation Closed-Loop for Model Capability Enhancement

Model capability is the central variable in LLM pre-training, yet is never observed directly: data shapes it prospectively, while evaluation reveals it only retrospectively, compressing samples, prompts, decoding, and scoring rules into one noisy score. Practical optimization runs this backward: a failure is observed first, and the engineer must infer the corpus fix. The two sides speak incompatible vocabularies -- benchmark names and per-sample correctness versus data sources, domains, and quality labels -- so this inference is usually intuition, not method. We close this gap with the \emph{capability slice}: a group of evaluation samples sharing background condition, task type, solving operation, and output constraint -- precise enough to localize a single weakness yet stable enough to survive aggregation, unlike a benchmark name, too coarse, or a single sample, too noisy. Built around this unit, an evaluation taxonomy, a non-instruction data taxonomy, and mapping rules form a closed loop turning a benchmark-level failure into a targeted, testable data intervention. We test this loop on two case studies pulling in opposite directions. First, the loop rules the data out: continued pre-training drives BBH down by $-46.82\%$, but diagnosis traces this to a single masked \texttt{\textless EOS\textgreater} loss rather than weakened reasoning; restoring it recovers BBH to $66.44$, above the original checkpoint, without changing the data. Second, the loop rules the data in: a persistent math-reasoning weakness is decomposed by solving operation into specific failing combinations, and a weakness-targeted sampling procedure built from it lifts AIME2025/AIME2026 Pass@128 from $6.67$/$0.00$ to $26.67$ each. The same unmodified loop reaches opposite, correct verdicts in both cases, showing the evaluation-to-data inference can be routine, auditable, and experimentally validated rather than intuitive.

GPTNT: Benchmarking Real-Time Collaboration Between Multimodal Agents on Keep Talking And Nobody Explodes

Multimodal models are increasingly deployed to solve tasks collaboratively with humans or other artificial agents. Existing benchmarks show that these models possess many of the required component capabilities, but the conditions that coincide in collaboration, including time pressure, information asymmetry, and imperfect communication, are usually studied in isolation. We introduce GPTNT, a benchmark built on the cooperative video game Keep Talking and Nobody Explodes, in which two agents must coordinate to defuse procedurally generated bomb puzzles against a live countdown. One agent can see and manipulate the bomb but does not have the defusal instructions; the other has the instructions but cannot see or manipulate the bomb. Neither agent can succeed alone: success requires effective and efficient communication. Unlike turn-based proxies, GPTNT requires agents to act asynchronously and communicate in real time. GPTNT is designed to separate collaboration from reliance on memorized solutions: the instruction manual, the partner, or both can be withheld to isolate what a model derives in the moment from what it already knows. We show that GPTNT poses a substantial challenge for state-of-the-art systems: none of the closed- or open-source models we test defuses a single bomb in real time, a bar that human players clear. Through controlled experiments, we identify critical weaknesses in state tracking, efficient action under time pressure, ambiguity handling, and error recovery. We release GPTNT as a benchmark for collaborative performance that current evaluations leave unmeasured. Because it runs on the real game, GPTNT benefits from procedural generation and inherits a living modding community, allowing the benchmark to evolve as models improve rather than being solved once and retired.

IMCBench: A benchmark for multimodal LLMs in Image-grounded Medical Conversations

Recent advances in large language models and vision-language models have enabled reasoning over multimodal data, offering opportunities for clinical applications such as decision support and triaging. However, existing medical AI benchmarks are fragmented: some support multi-turn dialogues but lack images, while others provide multimodal inputs but focus on single-turn QA tasks. To address this gap, we introduce IMCBench, an image-grounded, multi-turn medical conversation benchmark that pairs real, publicly available clinical images with synthetic patient profiles to simulate realistic patient-clinician interactions. Each conversation is evaluated across three clinical dimensions: safety, accuracy, and appropriate use of uncertainty in diagnosis. We benchmark eight multimodal frontier models across four model families (Claude, GPT, Nova, and Llama), scoring each on a 1-5 scale using LLM-as-Jury scoring calibrated against expert clinician annotations. Our results show that Claude Opus 4.6 achieves the highest overall score (3.61), followed by Claude Sonnet 4.6 (3.30) and GPT-5.2 (3.29), though no model dominates all dimensions and safety degrades for both malignant and rare conditions ($\Delta$ = -0.27 each). Ablation studies further reveal that both visual input and EHR context contribute to safe guidance (safety drops of 0.18 and 0.23 on average when each is removed), with stronger models leveraging visual features more effectively. Together, these findings demonstrate that accurate clinical description does not guarantee safe patient guidance, motivating the need for multi-dimensional evaluation frameworks in medical AI.

Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories

Current approaches to enhance Large Language Model (LLM) reasoning, such as Chain-of-Thought and "Wait" prompts, primarily encourage models to think more, yet often fail to guide them toward Truth. While Representation Editing (RepE) offers a intrinsic control, its application to dynamic reasoning trajectories remains underexplored. In this work, we bridge this gap by investigating the geometry of truth within unfolding reasoning chains. We uncover three critical insights: (1) Truth is encoded at the sentence level and is entangled with latent reasoning patterns; (2) Effective intervention follows an Uncertainty Principle and a Decay Effect, requiring localization to early, high-entropy forks; (3) Naive steering vectors suffer from noise, risking collateral damage to correct trajectories. Based on these findings, we propose DynaSteer, a dynamic RepE framework. DynaSteer employs pattern clustering to disentangle reasoning manifolds and utilizes Fisher-LDA to project purified truth. By dynamically monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary. Comprehensive experimental results on several MATH benchmark verify the effectiveness of DynaSteer, and experiments on out-of-domain coding tasks further confirm its generalization ability. Our code is publicly available at https://github.com/tianlwang/DynaSteer.

Aristotelian Virtue Profiling of LLMs through Ethical Dilemmas

Large Language Models (LLMs) often face ethical tradeoffs in which several responses may be defensible but express different priorities, such as fairness, honesty, courage, or restraint. We introduce VirtueMap, a framework for describing these patterns through an Aristotelian virtue-ethics lens. Instead of asking for a single correct answer, VirtueMap asks humans or LLMs to rank all five responses to each of seven general, non-lethal, non-political, and non-religious ethical dilemmas. To define the reference orderings used for scoring, we first proposed, for each dilemma and virtue, an ordering of the five responses from most to least expressive of that virtue. We then collected more than 100 respondent evaluations per ordering and retained it as operational ground truth only when at least 95% confirmed it. Rankings are scored against these retained orderings using normalized Borda alignment, yielding profiles over Practical Wisdom, Justice, Truthfulness, Courage, and Temperance. We apply VirtueMap to nine LLM families in a repeated-run evaluation and find high mean rank consistency (90.3%), with the largest differences appearing on Courage, Temperance, and Justice. We also release an interactive website that computes profiles locally in the browser and compares respondents with measured LLM profiles.

An AI agent for treatment reasoning over a biomedical tool universe

Treatment reasoning underpins every therapeutic decision, integrating disease context, comorbidities, medications, contraindications, and evolving biomedical knowledge to select an appropriate therapy. It is inherently iterative: candidates are weighed against many constraints, revised as evidence emerges, and grounded in verifiable sources. Here we introduce ATHENA-R1, an AI agent for treatment reasoning across all FDA approved drugs since 1939, trained by reinforcement learning over a universe of 212 biomedical tools. At each step it identifies missing information, selects and runs relevant tools, and incorporates the evidence. To train it without human-annotated traces, we build a two-level self-learning framework: multi-agent systems construct the tools, tasks, and reasoning trajectories for supervised fine-tuning, then reinforcement learning with scientific feedback rewards reasoning quality (evidence gathering, grounded tool use, logical non-redundancy). Across five benchmarks of 3,168 drug reasoning tasks and 456 patient treatment cases, ATHENA-R1 outperforms language models and tool-use systems, reaching 94.7% accuracy on open-ended drug reasoning and 82.9% on treatment reasoning, 17.8 and 10.7 points above GPT-5. In blinded evaluations by experts from 28 rare disease organizations, it is preferred over reference models on all criteria, and physicians rated it favorably on complex hospitalized cardiovascular and infectious-disease cases. Adverse-event hypotheses it generated, tested in electronic health records from 5.4 million patients, reached adjusted odds ratios of 1.48-1.84, with no elevation among negative controls. Because it requires knowing what evidence to seek before concluding, treatment reasoning has long been hard for AI; we show it can be reframed as a learnable process of iterative evidence gathering that reinforcement learning can train AI to perform.

COMPASS: Grounding Composition-Intent Guidance in Unified Multimodal Models

Composition is a high-level visual intent that governs where subjects are placed and how a scene is organized, yet current unified multimodal models remain unreliable at fine-grained composition recognition and struggle to turn such intent into controllable generation. We present COMPASS, the first unified multimodal framework that grounds composition-intent control in a single system spanning both composition perception and composition-guided generation, with a shared expert token $\tau_c$ as the central intent anchor. On the perception side, COMPASS injects composition expertise into an MoE backbone in a minimally invasive manner and distills the inferred intent into $\tau_c$. On the generation side, COMPASS reuses $\tau_c$ as a global conditioning signal that steers the denoising trajectory, effectively converting passive composition analysis into explicit layout control. To support systematic instruction-following composition learning and evaluation at scale, we construct Comp-11, a large-scale dataset with an 11-class taxonomy and reasoning-augmented annotations. Extensive experiments show that COMPASS substantially improves category-level composition understanding and delivers more composition-consistent, prompt-faithful generation than strong baselines.

BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards

Critic-free reinforcement learning with verifiable rewards (RLVR), exemplified by Group Relative Policy Optimization (GRPO), avoids training a value function (critic) and reduces memory and compute overhead relative to critic-based PPO pipelines for aligning large language models. However, GRPO-style advantage estimation depends on prompt-local (within-prompt-group) reward statistics and can be unstable. In particular, when all rollouts in a prompt group receive identical rewards, the within-group reward variance becomes zero, and group normalization yields zero advantages for that group, impeding learning in cold-start regimes with binary verifiers. We introduce BV-Blend, a critic-free framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. BV-Blend maintains EMA-tracked reward moments for each cluster, derives a confidence weight from a standard error of the mean (SEM) proxy, and uses this weight to blend historical and prompt-local baseline and variance statistics into a standardized advantage for PPO-style clipped updates. Experiments on verifiable reasoning benchmarks show that BV-Blend improves training stability and performance, and remains robust in regimes where group-normalized methods may stall.

The Two Genie Game: Adoption and Welfare in Audit-Grounded AI Governance

We ask under what conditions an agent with a harm-minimizing policy can displace an approval-seeking (RLHF) agent in a competitive market, and when that policy is sufficient to prevent community harm. We use evolutionary game theory (finite-population Moran-Fermi pairwise comparison) to formalize this subject to assumptions of wisher hindsight, peer testimony, a monotone harm ledger, sufficient information density of community feedback, and a finite, depleting resource pool, in a negative-sum environment. We show that adoption is favored when the prior distributions on how readily wishers attune to community sentiment are monotone, exhibit endpoint inversion, and have a centro-symmetric pairing property, and demonstrate this with several long-tailed priors (Hill, Pareto, Lomax, Frechet). Where it is favored, a critical adoption level separates communities that drift back to the approval-seeking agent from those for which the audited agent fixes; above that level fixation is the overwhelmingly likely outcome. We derive when fixation is attainable as a bound on the effective (informational) size N_c of the community, which must be small enough to allow fixation before depletion. We present these as Theorems 5.4 and 5.5; the algebraic and finite-grid backbone is machine-checked in Lean 4, with the barrier-crossing asymptotics retained as explicit hypotheses. We show that a self-audited agent with a community ledger is not, in general, sufficient to prevent community harm. Sufficiency depends both upon the alignment of the agent's audit with community values and the timeframe over which harm is evaluated. Regardless of alignment, once adoption reaches dominance, the state is absorbing. The same policy that reduced harm under alignment becomes a trap, welfare-negative under misalignment and, even under alignment, one that locks in harm deferred past the adoption horizon.

TrajRS: Towards Certified Robustness in Pedestrian Trajectory Prediction

The robustness of trajectory prediction models is crucial for developing safe autonomous driving systems. Adversarial attacks on trajectory prediction can significantly impair the accuracy of predicted trajectories, leading to hazardous driving behaviors. While heuristic defense strategies have been implemented to enhance the robustness of trajectory prediction models, these measures often fail against more sophisticated, targeted adversarial attacks. Hence, there is a pressing need to establish verifiable safety assurances for trajectory prediction models. In this paper, we extend the traditional Randomized Smoothing framework to "TrajRS", which provides a certified robust radius for smoothed trajectory predictors. We clarify and expand the formal definitions of robustness in trajectory prediction and tailor the practical TrajRS scheme specifically to "robustness for the optimal prediction" and "robustness for all possible predictions". An extensive set of experiments demonstrates that TrajRS effectively achieves robustness certification for all smoothed pedestrian trajectory predictors in this work.

ComMem: Complementary Memory Systems for Test-Time Adaptation of Vision-Language Models

Test-time adaptation (TTA) of vision-language models (VLMs) is essential for their robust deployment in dynamic, real-world environments. However, existing TTA methods often adapt locally without accumulating knowledge over time, or operating within a single modality without exploiting VLMs' inherently multi-modal nature. Inspired by the \textbf{Com}plementary \textbf{Mem}ory systems of the biological brain, we propose \textbf{ComMem}, an innovative approach that mimics the distinct but cooperative roles of the hippocampus and neocortex to enable effective TTA for VLMs. ComMem consists of two key components: a fast-adapting detailed memory, akin to the hippocampus, that forms a dynamic visual cache from high-confidence test samples; and a slow-integrating abstract memory, akin to the neocortex, that continually refines global textual prototypes. For each test instance, ComMem jointly optimizes both memory systems to ensure cross-modal consistency. Extensive experiments on 15 benchmark datasets show that ComMem significantly outperforms state-of-the-art methods under both natural distribution shifts and cross-dataset generalization, offering a promising direction for enhancing VLMs' practical adaptability.

Agentic Abstention: Do Agents Know When to Stop Instead of Act?

LLM agents are expected to act over multiple turns, using search, browsing interfaces, and terminal tools to complete user goals. Yet not every goal is well specified or achievable in the available environment. In such cases, a reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls. We define Agentic Abstention, the problem of deciding when an agent should stop acting under uncertainty. Unlike standard LLM abstention, which is usually evaluated as a single-turn answer-or-abstain decision, agentic abstention is a sequential decision problem: an agent can answer, abstain, or gather more information at each turn, and the need to abstain may only become clear after interacting with the environment. We study this problem across web shopping, terminal environments, and question answering, evaluating 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 tasks. Our results show that the main challenge is not only whether agents can abstain, but also when they abstain. Some agents never abstain when they should, while others do so only after many unnecessary interactions. This gap is especially large on tasks where the instruction appears feasible until the environment reveals otherwise (e.g., no valid result matches the instruction). We further find that model scale, reasoning, and agent scaffolding affect abstention in different ways, where larger or more capable models sometimes perform worse at timely abstention. Finally, we introduce CONVOLVE, a context engineering method for improving agentic abstention that distills full interaction trajectories into reusable stopping rules. On WebShop, CONVOLVE substantially improves timely abstention without updating model parameters, raising Llama-3.3-70B's timely recall rate from 26.7 to 57.4. Our dataset and code are available at https://lhannnn.github.io/agentic-abstention

Agent Safety Is Action Alignment

Large language models increasingly act as agents: they call tools, move money, delete records, and send messages on a user's behalf. To keep them safe, practitioners imported the chatbot-era recipe (train the model to refuse unsafe inputs) into the agentic setting, and treat the resulting capability loss as a manageable ``alignment tax.'' We argue this is a \emph{category error}. Refusal is a primitive for \emph{content safety}, where the harm is in the model's output and is therefore a learnable function of it. Agentic harm is different in kind: it lies not in any output but in the relation between the authority an action exercises and the authority the user granted, which is absent from the text the model sees. Importing content-safety methods into this regime does not trade capability for safety; it pays capability and buys negative security. We support this with three lines of evidence spanning the autonomy spectrum: defense-trained models learn surface patterns rather than intent; the same training collapses multi-step agents before any threat appears while leaving them exploitable; and even undefended frontier models exceed granted authority under ordinary use. We conclude that action safety cannot be installed in weights. It must be expressed as \emph{least privilege}, enforced \emph{outside} the model at the action boundary, and evaluated as \emph{action alignment} (a relational, deployment-conditioned property) rather than a refusal score.

Self-Supervised Theorem Discovery in a Formal Axiomatic System

Recent artificial intelligence (AI) systems have shown remarkable progress in mathematical reasoning. Many existing approaches, including large language models (LLMs), draw on human prior knowledge in the form of mathematical text, code, or theorem libraries. Although these approaches are highly effective in practice, it remains an open question whether an agent can autonomously discover useful theorems without such human priors. We study this question in a formal axiomatic system by developing an agent that starts from axioms and inference rules alone and gradually grows a library of useful theorems. Concretely, we propose a self-supervised theorem-discovery algorithm that alternates between proof search and useful-theorem extraction, building a theorem library whose entries are reused as lemmas for subsequent proof search. Experiments show that the agent discovers tens of thousands of theorems and finds proofs for human-written benchmark problems, suggesting that its discoveries include theorems meaningful from a human mathematical perspective. Furthermore, the discovered theorems improve LLM proof performance when provided as prompt lemmas, indicating that they can serve as external knowledge for LLM reasoning. Our results provide evidence that useful theorems can emerge from proof search without relying on human-provided theorem libraries. More broadly, they suggest a path toward self-evolving AI systems for mathematics whose discoveries remain formally verifiable.

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.

HyphaeDB: A Living Knowledge Topology for Agent-First Memory

Every existing vector database and agent memory framework treats memory as passive storage that agents query explicitly. No system propagates knowledge between agents through the memory layer itself. We introduce HyphaeDB, an agent-native memory infrastructure that reinterprets the Hierarchical Navigable Small World (HNSW) graph topology the data structure at the core of every modern vector database not as a search optimization, but as a communication fabric for multi-agent AI systems. In HyphaeDB, agents are nodes in the vector space with persistent positions, knowledge propagates via a gossip protocol through the graph's neighbor structure with energy-based attenuation, and emergent behaviors contradiction detection, pattern crystallization, and consensus formation arise from the combination of topology, propagation dynamics, and local interaction rules. We present the architecture built on three primitives (knowledge nodes, topology edges, and memory diffs), a multi-layer abstraction hierarchy with promotion via emergent consensus, and theoretical analysis grounding the system in small-world network theory, epidemic broadcast protocols, and swarm intelligence. We provide a reference implementation on PostgreSQL with pgvector and describe a concrete deployment in Swarm-Driven Development, a multi-agent software engineering methodology. HyphaeDB represents, to our knowledge, the first system to combine navigable small world topology with gossip-based knowledge propagation for multi-agent coordination.

Primary ICD Category Prediction using LLM-based Probing

Objective: ICD codes are central to reimbursement, research, and population health surveillance, yet automated coding systems often struggle to integrate diagnostic signals from both clinical narratives and structured electronic health record (EHR) variables. We evaluated whether frozen medical large language model (LLM) representations can serve as a shared embedding space for multimodal primary diagnosis category prediction. Materials and Methods: We constructed a MIMIC-IV cohort of 13,645 admissions from the 10 most frequent primary ICD-10 codes, consolidated into seven categories. Structured variables were serialized into clinical narratives and combined with leakage-pruned discharge notes. Using a frozen MedFound-Llama3-8B-finetuned backbone, we extracted hidden states from five transformer layers and trained linear probes for structured-only, unstructured-only, and combined inputs, comparing against XGBoost and information-matched PLM-ICD baselines and evaluating MIMIC-III adaptation with a compact bottleneck adapter. Results: The combined probe performed best on MIMIC-IV (87.69% strict; 91.45% medical accuracy), exceeding both single-modality probes and baselines. The structured-only probe outperformed its standard baseline by 6.19 points in medical accuracy. Diagnostic information became increasingly linearly separable in deeper layers, and a 2M-parameter adapter restored cross-dataset transfer to MIMIC-III using only 5% of target labels. Discussion: LLM embeddings can unify structured and narrative EHR information for multimodal diagnosis prediction, supporting efficient reuse of clinical representations across modalities and datasets through a small representation-level module. Conclusion: Multimodal probing of frozen medical LLM representations provides a practical approach for studying EHR modalities and adapting clinical representations across datasets.

MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes

Doctor agents are moving beyond single-turn answer generation toward evolving clinical decision systems. Within an outpatient episode, they acquire evidence, use examination and consultation resources, and decide when to finalize a diagnosis and management plan. Across episodes, their behavior may change through memory, retrieval, reflection, or other update mechanisms. Current evaluations only partially cover this setting. Fixed-input medical QA benchmarks score final answers from complete inputs, whereas many interactive benchmarks still focus on individual encounters or fixed runs, providing limited support for evaluating how episode-level decisions interact with cross-episode experience. We introduce MedEvoEval, an executable longitudinal evaluation framework based on action-gated simulated outpatient episodes. Each source case is converted into role-specific patient, examination, and manager views; evidence is revealed only through valid actions; and each episode records a structured trace that links observations, actions, final outputs, manager scores, and optional experience write-back. We release a runnable E&D artifact with 700 processed episodes, provenance notes, schemas, an episode runner, scoring scripts, configurations, example logs, analysis code, and trajectory- and step-level derivatives. Experiments show that episode traces expose process costs hidden by final-answer scoring, show how MDT-style consultation reallocates resources, and support longitudinal analyses of memory maturation, held-out transfer, update-stage response, and backward retention. Together, these results show that MedEvoEval provides a concrete basis for evaluating whether doctor agents improve through experience, transfer useful behavior, and retain earlier capabilities over time.

Expert Evaluation of Clinical AI Tools on Real Point-of-Care Clinical Queries

Physicians now pose millions of clinical questions to AI tools each week, yet these tools are evaluated largely on hypothetical or exam-style questions, not those actually asked in practice. We report a blinded evaluation built on 620 Real-world Point-Of-Care Queries (Real-POCQi) submitted to the OpenEvidence (OE) platform by physicians spanning 30 specialties, as well as 187 questions from HealthBench. 149 practicing physicians across 36 states made head-to-head comparisons between answers from three frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5) and a specialized clinical tool (OE), with graders matched to each question's specialty. When comparing answers along five dimensions relevant to clinical decision support -- accuracy, clinical utility, source quality, verifiability, & completeness -- physicians scored the specialized tool highest on all axes; in the primary analysis on Real-POCQi, win differences (margins between win and loss rates) ranged from 25 to 39 percentage points (p<0.001). Results remained consistent in sensitivity analyses stratifying by citation display, answer length, OE-user status, and Real-POCQi versus HealthBench. In parallel, LLM judges were found to systematically differ from expert judges, though both generally agreed on the best model. These findings underscore two conclusions: (i) AI tool evaluations should reflect real-world query distributions and use expert judges that mirror the specialization defining modern medicine and (ii) the consistent advantage of the specialized tool over general-purpose models does not necessarily mean that the latter cannot serve similar purposes, but that targeted engineering and customization can yield meaningful gains in performance for its users. We release Real-POCQi as a public benchmark, as well as the prespecified statistical analysis for reproducing results of this study.

Customized Generative AI Agent for Transportation Engineering Practice: A Development and Continued Pre-training Guideline

Recent advancements in generative artificial intelligence (AI) and large language models (LLMs) have shown significant promise in automating complex reasoning, summarization, and question-answering tasks. However, the effectiveness of general-purpose LLMs in specialized engineering domains remains limited due to insufficient exposure to technical standards, engineering terminology, and domain-specific semantics. This study proposes a systematic approach to developing a customized generative AI agent for transportation engineering applications. A curated corpus of U.S. transportation manuals, design guidelines, and regulatory documents is used to conduct continued pretraining of six state-of-the-art LLMs through a unified low-rank adaptation (LoRA) framework. The training process is monitored to ensure convergence and model stability. Performance is evaluated using standard natural language processing metrics, including BLEU-4 and ROUGE, with Qwen2.5-7B and LLaMA-3.1-8B demonstrating the highest domain alignment and response quality. Results validate the effectiveness of LoRA-based adaptation in improving LLM performance on technical content interpretation and context-specific reasoning. This work contributes a reproducible development framework for constructing domain-specialized generative AI agents, supporting broader deployment in transportation research, design, planning, and policy analysis.

Preventing Error Propagation in Multi-Agent AI through Runtime Monitoring

Multi-agent AI systems can improve answer selection by allowing different language models to exchange reasoning traces, revise initial predictions, and support a final decision. However, such communication may also introduce reliability risks: reasoning from one agent can correct another agent's mistake, but it can also mislead an agent that was initially correct. This paper studies reliable multi-agent AI communication through reasoning exchange and runtime answer revision. We develop a framework in which agents first answer multiple-choice questions independently, then share reasoning traces and revise their decisions. We conduct numerical experiments where we evaluate whether this process improves accuracy, produces more positive than negative answer transitions, and remains effective across domains such as cybersecurity, networking, and general knowledge. The results help identify when multi-agent reasoning improves reliability and when it may propagate errors.

Memory as an Attack Surface in LLM Agents: A Study on Multiple-Choice Question Answering

AI agents extend conventional large language model (LLM) applications by integrating language understanding with task execution, external tool use, and memory mechanisms. While memory allows agents to retain prior interactions and provide more personalized and context-aware responses, it also introduces a new vulnerability: information stored in memory can influence future outputs even when the current query is clean. In this paper, we investigate memory manipulation in LLM-based agents for multiple-choice question answering. We first design and implement an LLM-based AI agent with an external memory component that stores and retrieves task-relevant information. We then introduce basic memory manipulation scenarios in which misleading or corrupted memories are inserted into the agent before it answers multiple-choice questions. Using a controlled experimental setup, we compare the agent's performance before and after memory manipulation and measure changes in answer accuracy, attack success rate, and selection of manipulated options. Our results show that even simple memory manipulations can noticeably affect the agent's final answers, causing it to select incorrect options despite receiving clean and well-formed questions.

Low-cost concept-based localized explanations: How far can we get with training-free approaches?

Concept-based Explainable AI (C-XAI) seeks human-understandable explanations grounded in semantic concepts, yet validation is limited by the scarcity of fine-grained concept annotations. We evaluate whether mid-scale Multimodal Large Language Models (MLLMs) can perform localized concept naming under strict zero-shot conditions by assigning labels to bounding-box regions at both object and part levels. We propose a reproducible zero-shot evaluation protocol for Concept Naming (CoNa) with (i) closed-set, category-constrained prompting for moderate vocabularies and (ii) Open-CoNa, an embedding-similarity-based strategy for large label spaces. Experiments with four MLLMs (7B-32B) show consistent performance trends across datasets, reaching 62%-88% object-level exact-match accuracy, highlighting the potential of training-free concept annotation from localized regions. We discuss limitations and failure modes and release a reproducible framework to support future low-cost C-XAI research.

Managing the Human Fallback: Skill Investment Under Improving AI and Worker Mobility

When firms deploy autonomous AI, they must decide how much work to leave to the system and how much to keep workers engaged. This decision affects current output and future human capital. We develop a parsimonious two-period model in which AI may outperform the worker when it functions, but may fail with positive probability. A firm chooses worker engagement; engagement lowers current output for below-benchmark workers, but changes future skill through learning and erosion. We distinguish two dimensions of AI progress: capability, the system's output when it works, and reliability, the probability that it works. In a single-firm benchmark, engagement is valuable only as fallback investment. The firm engages the least-skilled workers most, because they have the largest skill gaps and are least costly to bring toward a useful fallback level. With worker mobility, engagement also affects labor-market sorting: workers prefer jobs that build more valuable skill trajectories. This sorting motive targets higher-skill workers near the AI frontier, where skill gains are more valuable and engagement is less costly. Mobility can therefore reverse the engagement pattern, shifting investment from the least-skilled toward the most-skilled workers below the AI benchmark. Mobility also reshapes how AI progress affects engagement: greater capability raises engagement by increasing the value of the skill trajectory a firm offers, whereas greater reliability can raise or lower it because it reduces fallback need while also changing learning opportunities. Under worker mobility, human-AI work design becomes a problem of human-capital investment, in which allocating work today shapes future skill.

Characterizing Large Language Model Agentic Workflows: A Study on N8n Ecosystem

Large Language Models (LLMs) are rapidly being adopted in low-code and no-code automation platforms, where non-expert users design workflows that combine natural language understanding with external services and APIs. LLM agents are LLM systems that use LLMs as a core "brain" to reason, plan, and autonomously execute complex, multi-step tasks. In this paper, we present the first large-scale empirical study of LLM agentic workflows in low-code automation platforms. We analyze more than 6,000 publicly available n8n workflows and examine four aspects of their design: task distribution, structural and tool use patterns, reliability mechanisms, and autonomy levels. Our analysis shows that LLM workflows are not merely prompt response pipelines. Instead, LLMs are commonly embedded within broader automation structures involving control logic, external tools, communication services, storage systems, and human review points. We further find that while many workflows include lightweight post-processing or routing logic after LLM execution, explicit reliability mechanisms such as structured fallback paths, repair loops, failure-specific alerts, and human approval gates remain relatively uncommon. These results reveal a gap between the increasing deployment of LLM agents in practical automation ecosystems and the limited engineering support for reliability, safety, and governance. Overall, our study provides ten empirical findings and five research takeaways for researchers, platform developers, and practitioners seeking to understand and improve real-world LLM agentic workflows.

HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) often relies on communication to mitigate partial observability, yet most existing protocols treat messages as flat dense vectors detached from the structure of the observations they summarize. This design overlooks an important source of inductive bias in many cooperative environments, where observations naturally follow a hierarchy such as groups and entities. We propose \textsc{HiComm}, a plug-in communication module that grounds messages in the sender's hierarchical observation. \textsc{HiComm} is receiver-driven: the receiver issues a query, and the hierarchy is resolved through a three-stage decoding process that first selects a group, then a sender, and then an entity within that group, returning the corresponding feature slice as the message. This converts communication from unstructured vector transmission into structured information retrieval over the sender's observation hierarchy. We instantiate this mechanism with Straight-Through Gumbel-Softmax for differentiable discrete selection and a lightweight shared projection design that attaches to standard MARL pipelines. Experiments across cooperative MARL tasks with different observation structures and coordination demands show that \textsc{HiComm} matches or outperforms representative learned communication baselines while reducing communication volume by up to $23\times$ per receiver per episode.

Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement

Discrete flow models have recently shown promising performance on few-step text generation; however, when naively applied to structured reasoning tasks such as Sudoku and Zebra puzzles, they converge confidently to incorrect answers (solving only $\sim$36% of Sudoku puzzles). We introduce Flow Reasoning Models (FRMs), a training and test-time-scaling framework for structured reasoning with flow models. We make the observation that, despite their poor solve rate, flow models can act as their own verifiers. A correct answer is a stable fixed point of the denoising dynamics, returning to itself when re-noised and re-solved. This enables a test-time-scaling paradigm: propose many candidate solutions and keep those that are dynamically stable, which alone reaches high solve rates on Sudoku-Shah (~$100\%$) and Zebra ($95.9\%$). This even generalizes to harder out-of-distribution puzzles like Sudoku-Extreme ($96.1\%$), without ever training on that distribution. This pure search, however, wastes a great deal of computation generating incorrect candidate solutions. We therefore design a training recipe to improve the base model's efficiency. First, we train flow models with a self-conditioning channel and close it at inference, letting them refine their own past predictions. Second, we train models to avoid their own failed generations using direct preference optimization. These changes substantially improve the base model's efficiency, letting it reach $99.2\%$ on Sudoku in just $7$ forward passes, over $8\times$ fewer than the strongest matched masked-diffusion baseline we compare needs for the same accuracy. When combined with test-time scaling, this lets flow models solve hard out-of-distribution puzzles (e.g. Sudoku-Extreme) far more efficiently.

Pooled Leaderboards Hide System-Specific Winners: A Reporting-Protocol Audit of Offline Root-Cause Analysis Benchmarks

Offline root-cause-analysis (RCA) benchmarks commonly rank methods by a single pooled top-1 accuracy across multiple subsystems, and engineers often read the pooled winner as a recommendation for their own subsystem. We audit that reading on three public RCA benchmark families -- OpenRCA, RCAEval, and PetShop -- covering 11 subsystems and 778 matched scoring units. To keep pairwise comparisons on identical cases, the main analysis retains four methods or comparators with complete coverage: BARO, a CD-1min adapter, max-$|Z|$, and per-service alert-count. All six pairwise comparisons show subsystem-level effects of both signs, every random-effects 95\% prediction interval crosses zero, and case-level interaction tests reject exchangeability in 5 of 6 pairs. Leave-one-system-out selection picks the lower-scoring method on up to 5 of 11 held-out subsystems, with regret reaching 24.8 pp on RCAEval / Sock-Shop. We release a 320-line audit module; given a matched RCA benchmark score table, it recomputes the same per-subsystem stability checks alongside pooled scores.

Direct Causation in International Humanitarian Law and the Challenge of AI-Mediated Civilian Cyber Operations

International humanitarian law protects civilians from direct attack unless and for such time as they take direct part in hostilities, with the ICRC's 2009 Interpretive Guidance operationalising this rule through a three-criterion cumulative test. This paper argues that AI-mediated civilian cyber operations challenge the direct causation element of this test in a structurally specific way: when a civilian deploys an autonomous multi-agent cyber system of the kind recently demonstrated in offensive AI research, the "one causal step" standard fails because harm is produced by system-generated decisions made after human disengagement, and the integral-part requirement does not extend because it presupposes downstream human contributors whose conduct can be independently classified. The framework therefore defaults to treating such deployments as indirect participation, in tension with its purpose of capturing civilians who personally take part in hostilities. Beyond the doctrinal analysis, this paper identifies goal-specification granularity as the property on which the integral-part test's concreteness component implicitly turns, classifies AI-mediated operations along a five-level spectrum, and argues that existing technical AI governance instruments do not log or report this property.

Selective Memory Retention for Long-Horizon LLM Agents

When does retention matter for memory-augmented LLM agents? We study this with TraceRetain, a lightweight framework for bounded external memory in frozen LLM agents that scores entries by interpretable features (success, age, access frequency, redundancy, specificity, similarity, downstream utility) and evicts the lowest-scoring ones at capacity. On clean ALFWorld with gpt-5-mini, external memory robustly improves over no memory across two seeds, but differences among bounded retention policies fall within Wilson 95% CIs: clean ALFWorld at T=100 to T=200 does not naturally exhibit the memory pollution retention is designed to address. Under a controlled noisy-write stress (75% synthetic distractors), unbounded memory and FIFO-K50 degrade on Precision@5 (20.2% to 12.4% and 15.8% to 3.8%) while TraceRetain-CEM is essentially unchanged (16.9% to 16.6%) and preserves 97/100 task success. The mechanism: unbounded memory has the highest mean similarity (0.87) but lowest precision, indicating failed distractors close to the query in embedding space. Held-out in-distribution evaluation shows memory-augmented policies solving 47 to 49 of 50 tasks vs. 39/50 for no memory. Bounded retention buys memory and step efficiency on saturated clean benchmarks at no task-success cost, and only differentiates from cache heuristics when streams contain noise.

Measuring Graph-to-Graph Semantic Similarity in Knowledge Graphs: An Empirical Evaluation of Knowledge Graph Embeddings

A Knowledge Graph (KG) represents facts as structured triples and is widely used to organize relational knowledge across diverse domains. Just as textual information ranges from words and sentences to complete documents, KG information can be interpreted at multiple levels, from entities, relations, and triples to subgraphs and entire KGs. However, existing KG embedding methods mainly focus on entities, relations, and triples, leaving graph-level semantics largely unaddressed. Conventional graph-level methods, which typically compare graphs based on structural patterns, are also insufficient because structural similarity alone cannot guarantee semantic similarity between KGs. To evaluate how well different methods capture such graph-level semantic information, we study graph-to-graph semantic similarity, which determines whether a pair of KGs represents semantically corresponding underlying information. To obtain reliable ground-truth correspondences, we construct a semantic matching dataset by modifying text documents, extracting KGs from both original and modified documents, and transferring their known correspondences to KG pairs. We compare text-based, structure-based, and KG embedding-based approaches on each dataset. For the KG embedding-based approach, we introduce two scoring functions: \textit{EmbPairSim}, which uses maximal pairwise entity similarity, and \textit{AvgEmbSim}, which uses a frequency-weighted centroid. Experiments on WikiText-2 and CC-News show that \textit{EmbPairSim} achieves up to 5.3 pp higher MRR than Sentence-BERT while using substantially fewer parameters. These results suggest that KGE representations can serve as compact and effective signals for graph-to-graph semantic similarity in KGs. Our code is available at https://github.com/SeungRyeolBaek/KG-to-KG-Semantic-Similarity.

Evidence-Informed LLM Beliefs for Continual Scientific Discovery

Open-ended scientific discovery with large language models (LLMs) increasingly operates as a long-horizon loop of hypothesis search and verification, where a reward signal guides which hypotheses to test next. A notable recent example is AutoDiscovery, which uses "Bayesian surprise" - the belief shift an LLM undergoes after observing evidence for a hypothesis - as both a discovery metric and a reward for search. We first observe that AutoDiscovery treats surprisal as a static quantity, while surprisal in human reasoning is non-stationary - it is defined relative to beliefs that evolve with experience, a prerequisite for continual scientific discovery. We address this mismatch with evidence-informed LLM beliefs: priors updated with evidence from previous hypotheses to compute non-stationary surprisal for new hypotheses. We compare in-context belief-updating mechanisms and find that embedding-based retrieval-augmented generation over prior discoveries best anticipates eventual posteriors, identifying 37.5% of static surprisals as spurious. We then modify search to avoid these spurious rewards and prioritize hypotheses that remain surprising under non-stationary beliefs. Concretely, we introduce two complementary changes to the original search procedure: belief-update filtering and diversity maximization. Across five discovery domains, our method increases accumulated non-stationary surprisal by 30.62% on average compared to the original search procedure, demonstrating that continual scientific discovery with LLMs requires not only better belief measurement but also search procedures that avoid redundancy and encourage diversity.

AI Trading's Alpha Singularity: Emergent Market Reasoning through Agent-to-Agent Self-Evolution

Automated alpha mining holds the scoring function fixed and varies the search algorithm over it. A search that converges against a fixed scorer overfits whatever the scorer cannot penalize, a primary cause of the out-of-sample generalization gap. We treat the scoring function as a search artifact alongside the alpha factors and study what conditions make this joint search admissible. Sealed Joint Search (SJS) is a framework: a set of structural conditions on information flow in an autonomous-discovery system that prevent joint search from collapsing into self-confirmation while keeping the evaluator sealed. Conditions cover role decomposition, typed inter-role communication, provenance-sealed reads, versioned stores, and substrate-local promotion. Agora tests SJS empirically: five LLM agent classes communicate via three channels, evolving eight skill libraries, with alpha libraries built on AlphaGen operators. Three evaluators write reports aggregated into one brief, carrying forward disagreement instead of voting. We run Agora for 100 rounds on CSI 1000 and evaluate on a 91-day 2026 holdout sealed from all LLM inputs. Agora achieves holdout Sharpe +1.87; best baseline +1.334 at favorable seed and -0.755 cross-seed mean. Pre-loading Agora's two metrics into a frozen-library ablation recovers only +0.40 of the +2.25 Sharpe gap, and adding PPO without library evolution worsens the gap. The two metrics emerge rather than being designed. Caveats: single-seed run, short-side concentrated signal, intended for long-short.

A Cognition-Emotion-Personality Framework for Modeling Human-Like Awareness and Behavior in Emergency Evacuations

Agent-based evacuation simulations are widely used to study crowd behavior during emergencies, but many models rely on assumptions such as perfect event awareness, complete exit knowledge, and fully rational decision-making. This paper presents an extended evacuation framework that integrates cognitive, emotional, social, and personality-related mechanisms into a unified model of human behavior under uncertainty. The framework incorporates a dynamic event-awareness mechanism based on a continuous Event Certainty Level, a memory-based representation of exit knowledge subject to acquisition, forgetting, and recall, a continuous fear model in which panic emerges as a high-intensity state, and an OCEAN-based personality representation. Neuroticism is explicitly integrated into the emotional model, influencing fear generation, escalation, social contagion, and recovery. Behavioral heterogeneity is further captured through individualized decision thresholds that affect responses to perceived risk. The framework is evaluated through simulation experiments examining the effects of spatial familiarity, memory robustness, decision sensitivity, emotional dynamics, and personality variation. Results show that cognitive, emotional, and personality-driven processes substantially influence evacuation dynamics, reducing evacuation efficiency and generating realistic crowd phenomena such as delays, confusion, injuries, and socially influenced behaviors. The proposed framework provides a more realistic representation of human behavior in emergency evacuations and supports systematic investigation of the interactions between cognition, emotion, personality, and crowd dynamics.

PolicyGuard: A Dialogue-Grounded Sub-Agent Verifier for Policy Adherence in LLM Agents

LLM agents handle user requests on behalf of organizations through tool calls and must follow the company policies stated in their system prompts. Prior work approaches this as a safeguarding problem -- external checks that block non-compliant agent actions. We argue that policy adherence is a broader problem: real workflows unfold across many turns, require explicit user confirmation and prerequisite reads, and hinge on the content of the dialogue rather than on any single argument value. Meeting this bar requires (i) full conversation context, (ii) self-reasoning over the policy and the current dialogue, and (iii) conversation-specific remediation that guides the agent's next turn -- three capabilities that prior safeguard work has often underestimated. We introduce POLICYGUARD, a sub-agent verifier that shares the agent's view of the dialogue, reasons over the policy in context, and provides actionable feedback for the agent's next turn. On tau^2-BENCH airline across three vendors (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro) with four trials per setting, POLICYGUARD improves PASS4 by +12.0 / +6.0 / +12.0 pp. Per-call analyses show POLICYGUARD achieves higher policy-violation recall while blocking roughly half as often as argument-level guards.

SurgVLA-Bench: Towards Evaluating Vision-Language-Action Models for Laparoscopic Surgical Robotics

Vision-Language-Action (VLA) models represent a promising direction for embodied intelligence in surgical robotics. Despite the prevalence of VLA benchmarks for general robotics, standardized evaluation platforms specifically designed for surgical contexts remain absent. To address this limitation, we present SurgVLA-Bench, the first comprehensive benchmark for evaluating VLA models in laparoscopic surgical robotics. Leveraging the SurRoL simulation platform, we construct a hierarchical task taxonomy ranging from atomic actions to complete surgical procedures, complemented by a multi-dimensional evaluation framework assessing action accuracy and semantic consistency. We then systematically evaluate two representative paradigms, including autoregressive models such as OpenVLA, and flow matching models such as $\pi_{0}$, $\pi_{0.5}$, and SmolVLA. Our experiments show that autoregressive models tend to excel in semantic understanding, while flow matching models often achieve higher task precision but may face generalization trade-offs. However, even the best-performing models remain far from satisfactory, as the constrained endoscopic field of view, restricted viewing angles, and frequent occlusions persist as fundamental physical bottlenecks. The code and data are available at https://github.com/VCL-HNU/SurgVLA

When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysis

Financial decision-makers face more information than they can directly inspect, making context compression necessary. Yet when large language models (LLMs) compress financial source material, they can alter the investment judgment supported by the original source. We frame this problem as information fidelity: compression loses fidelity when it changes the decision induced by the source. In agentic systems, such losses may recur across intermediate steps and amplify throughout the decision process. Across financial filings and earnings-call transcripts, we find that LLM-based compression can produce fluent and factually plausible compressed contexts that nevertheless alter downstream decisions. We analyze two diagnostic patterns associated with fidelity loss: decontextualization, where salient evidence is retained but separated from the caveats and contextual qualifiers needed for correct interpretation, and model dependency, where different compressors expose different views of the same source. We then propose Agentic Context Compression, which generates multiple candidate compressions and audits their disagreements against the original source. Our results suggest that financial compression should be evaluated not only by efficiency or factuality, but also by its ability to preserve decision-relevant context.

The Complexity Ceiling Benchmark: A Multi-Domain Evaluation of Sequential Reasoning Under Depth Scaling

We introduce the Complexity Ceiling Benchmark (CCB), a controlled evaluation of how language-model reasoning decays as the number of required sequential steps grows. CCB fixes the semantic content of a task and varies only its depth N in {5,...,50} across three structurally distinct regimes: grounded spatial state-tracking, abstract symbolic pointer manipulation, and transitive relational inference. Across 6,000 trials over five frontier and open-weight LLMs we find a consistent pattern of geometric per-step decay with widely separated domain ceilings: on the first two regimes the strongest models retain pd>0.92 across N=50; on the third every model collapses by N=5, with the best model's 50%-success horizon at H0.5~4.7 steps despite pd=0.863. A trace-level metric (TFBC) shows that 14.5% of correct answers across the benchmark are reached via incorrect intermediate reasoning. Forced verbose state-tracking does not move the ceiling (McNemar p=1.000), and the mean step at which reasoning first diverges, k*, predicts within-domain accuracy better than parameter count. CCB and the geometric decay model together reduce a model's long-horizon reasoning profile to one interpretable number per task family.

Process Advantage Signal Shaping: A Paradigm-Agnostic Middleware for Process-Supervised RL in LLM Reasoners

Group Relative Policy Optimization (GRPO) is a default recipe for process-supervised reinforcement learning of LLM reasoners, and dense process supervision -- via learned process reward models (PRMs) or on-policy-distillation KL signals -- is a common way to densify its otherwise weak outcome reward. Layering such a step-level signal on top of GRPO's group-standardized advantage, however, exposes three structural pathologies: \emph{channel contamination} between the pooled process, outcome, and format streams at group standardization; \emph{resolution mismatch} between the granularity of the process signal and the granularity of the logical decisions being credited; and a \emph{cumulative trap} by which GRPO's return-to-go sum surfaces either length inflation or truncated exploration depending on the sign regime of the signal. We propose \textbf{PASS} (\emph{Process Advantage Signal Shaping}), a compact middleware that sits between any scalar step-level process signal and GRPO's clipped surrogate and addresses the three pathologies in turn: \emph{Advantage Fusion} standardizes the three streams independently within each group, \emph{Chunk-by-Value} derives value-homogeneous chunks from the signal itself and broadcasts credit within each chunk, and \emph{Divide-Length} converts the cumulative objective into an average-value-density score. We validate PASS across two domains and two process-signal paradigms -- a learned PRM on mathematical reasoning and an on-policy-distillation KL signal (with a generalized variant) on multi-hop question answering -- and under two group-standardization operators. In every regime PASS delivers a consistent pass@1 gain over the corresponding GRPO baseline.

Hierarchical Experimentalist Agents

Large language models (LLMs) are increasingly used to take actions in the real world and support human decision-making, yet most agents rely on parametric knowledge, fixed post-training data, retrieval, or search. This paradigm breaks down in novel domains and for sophisticated queries that cannot be answered from prior knowledge alone. Knowing the laws of physics, for instance, does not by itself enable LLMs to answer queries or complete long-horizon tasks in a complex physical system. To address this, we introduce Hierarchical Experimentalist Agents (HExA), an in-context self-improvement framework to learn from active experimentation. HExA iteratively designs and refines query-relevant experiments, learns a reusable library of composable skills from experience, and integrates experimental evidence to answer queries or take actions. HExA is training-free, compatible with any black-box model, and does not require external supervision, oracles, or offline data. To evaluate active experimentation, we introduce Interphyre, a tool-calling benchmark built on the PHYRE 2D procedural physics environment, where agents propose interventions and test hypotheses through simulation APIs. Experiments show that current LLM agents struggle in these settings, especially on the hardest levels of Interphyre. Claude Sonnet 4.6 achieves only 2% success, while HExA improves the same model to up to 77% success. HExA also improves open-weight models and outperforms agentic baselines such as ReAct and Reflexion. Moreover, using only skills learned from easier levels and transferred without active experimentation, HExA achieves 44% success, demonstrating the reusability and generalization of its learned skills. Overall, HExA shows that learning through active experimentation can help agents discover useful knowledge, acquire reusable skills, and make efficient progress on novel long-horizon tasks.

PHF: Privileged Hidden Flow for On-Policy Self-Distillation

On-policy self-distillation (OPSD) trains a reasoning model on rollouts sampled from its own policy by matching a privileged teacher that also sees verified reference solutions. Existing OPSD objectives supervise only the output distribution, so privileged context affects training through a token-level divergence without directly supervising the internal computation that produced that distribution. We propose Privileged Hidden Flow (PHF), which additionally distills how a privileged teacher's hidden states move along the same rollout. Rather than forcing each student hidden vector to match the teacher vector at the same token position, PHF aligns token-to-token transition directions and trajectory geometry over selected generated positions. The all-layer recipe also includes an adjacent-layer relation computed from these same transitions, without pointwise hidden-state imitation. Under the same 100-step training schedule, PHF improves the Average@12 aggregate over our reproduced OPSD baseline on Qwen3-1.7B, 4B, and 8B, with observed gains of about +2.2, +1.5, and +1.7 points. The transport objective is exactly invariant to shared trajectory offsets; its local geometry term is also invariant to orthogonal transformations of transition directions. Ablations distinguish the fixed PHF recipe from pointwise hidden-state matching, single-channel transition losses, and layer-subset choices, supporting PHF as a compact hidden-flow extension to OPSD.

When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning

Chain-of-Thought (CoT) improves large language models (LLMs) on difficult reasoning tasks, but it often incurs long natural-language rationales that are poorly aligned with efficient machine reasoning. We propose Communicative Language Symbolism Routing (CLSR), a test-time framework in which multiple LLM agents autonomously invent, evolve, and share compact Language Symbolism Frameworks (LSFs), while a latent-free router adaptively selects and composes these languages per query to optimize the accuracy-token trade-off. Unlike prompt optimization that refines surface instructions, CLSR treats each LSF as a reusable symbolic protocol with compact symbols, usage rules, and a message-passing contract, and improves it through an evolutionary loop driven by correctness and token cost. At inference time, the router may invoke a single low-cost LSF call, ensemble multiple LSFs, or execute a multi-round LSF composition protocol on harder queries. Across challenging benchmarks, CLSR reduces latency-oriented generated token completion by $3\sim 6\times$ compared to standard CoT while maintaining accuracy. We further derive an information-theoretic lower bound on token cost under arbitrary symbolism and show that, under an interpreter-realizability premise, multi-round LSF protocols conditionally subsume program-execution pipelines. Code is publicly available (https://github.com/pzqpzq/LSF_MDia).

Diagnosing and Repairing Factual Errors in RAG under Budget Constraints

Retrieval-Augmented Generation (RAG) improves the factuality of large language models by grounding responses in external evidence, yet real-world deployments remain fragile. Failures often stem from missing or weakly relevant evidence, as well as from generation that does not faithfully reflect the retrieved context. Many existing approaches rely on fine-tuning, privileged access to internal model signals, or resource-insensitive escalation strategies, which limits their practicality in black-box and budget-constrained settings. We propose D2R-RAG (Diagnose-to-Repair RAG), a model-agnostic and resource-aware framework that combines lightweight failure diagnosis with adaptive repair. D2R-RAG derives interpretable failure signatures from observable signals in the query, retrieved evidence, and generated response, and then selects from a small set of corrective actions under explicit latency and VRAM constraints. Experiments on FEVER and HotpotQA show that D2R-RAG improves reliability over recent baselines and achieves better accuracy--efficiency trade-offs across multiple compute budgets. The code is available at https://github.com/CyberScienceLab/D2R-RAG/.

LLM-Guided Planning for Multi-hop Reasoning over Multimodal Nuclear Regulatory Documents

Reviewing nuclear regulatory documents requires multi-hop reasoning across tens of thousands of pages, where judgments depend on evidence assembled across multiple chapters. We frame this task as planning: an LLM-based agent observes the evidence collected so far, picks the next document fragment to inspect, and stops when the evidence is sufficient. The agent operates over a vectorless document tree using browse, read, and search tools, and maintains a dynamic knowledge graph (KG) as state. On a 200-question benchmark over NuScale Final Safety Analysis Report (FSAR) documents, the system reaches 81.5% accuracy with a RAGAS Faithfulness of 0.93. The dominant performance factor is planning: against PageIndex, which uses the same document tree without state-conditioned action selection, the gap is +38.0pp (43.5% to 81.5%, p<0.001). The system also outperforms LightRAG (73.0%, p<0.05), HippoRAG (70.5%, p<0.01), and GraphRAG (49.5%, p<0.001), and matches RAPTOR (75.5%, p=0.11) without offline indexing. Edge inference adds 2.8x cost without raising accuracy; we retain it as a traceability module. Of 7,391 inferred edges, 3 Violates edges (0.04%) flag scope boundaries (Q058) and partial conformance (Q176) as typed annotations that a human reviewer can audit.

Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning

Existing multi-agent debate frameworks suffer from two critical limitations: they rely on static architectures where agent roles and coordination patterns are fixed at design time, and they require instantiating multiple model copies, incurring substantial computational overhead. We propose Mixture of Debaters (MoD), a unified framework that enables dynamic self-debate within a single model by leveraging the Mixture-of-Experts paradigm. We address three key challenges in adapting MoE for dialectical reasoning: (1) dual-routing that decouples role allocation from process flow, dynamically determining when to debate versus when to synthesize; (2) momentum switching that smooths token-level routing with local context, reducing expert-switch jitter; and (3) unified self-debate that encapsulates diverse debating personas into lightweight expert modules, eliminating inter-agent communication while preserving behavioral diversity. Extensive experiments on multimodal benchmarks demonstrate that MoD outperforms both single-model baselines and conventional multi-agent systems, achieving superior accuracy with 3.7x lower latency and 87% reduction in token consumption.The source code can be accessed at https://github.com/YongLD/MoD.

FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Models

Despite the impressive capabilities of Large Vision-Language Models (LVLMs), they remain susceptible to hallucination, generating content inconsistent with the input image. Recent studies attribute this to the dominance of language priors over visual inputs and employ contrastive decoding methods to mitigate this dominance, but the mechanistic origin remains unexplored. We investigate the information flow through each transformer layer and find that attention modules consistently aggregate visual evidence, while FFN modules at critical layers act as the source of language priors. These priors can override visual evidence, causing correct predictions in intermediate layers to drift toward incorrect outputs. Based on this insight, we propose FADE (FFN Attenuation for DEcoding), a training-free method that attenuates FFN outputs to reduce language-prior dominance. Evaluations on POPE, CHAIR, and MME benchmarks across LLaVA-1.5, mPLUG-Owl2, and InstructBLIP show that FADE effectively mitigates hallucinations while preserving inference efficiency.

How Much Due Diligence Before You Bid? Learning in Intractable Takeover Auctions

When two companies bid to buy the same target, no one knows exactly what the target is worth. Each bidder pays for due diligence: costly, imperfect homework that sharpens its own private estimate before it bids. How much of that homework is worth buying? We build a simple computer model of the bidding contest and let it teach itself to bid well by playing against itself, the way a game engine learns chess. The economic question, how much diligence pays for itself, and the computational question, when the contest becomes too complex to solve exactly, are both controlled by a single thing: how many pieces of private information a bidder carries. Our main finding is that the right amount of diligence is modest and finite. It falls as diligence gets more expensive, and it falls further when both sides are doing their homework, because competition erodes the value of knowing more. We also test a recent claim from AI research: that simple, general self-play methods can rival the specialized, expensive algorithms usually built for games like these. Running on an ordinary laptop with no costly frontier AI, we find the simple methods are the best of the self-learning approaches, though purpose-built exact methods still win whenever the game is small enough to solve outright. The simple methods earn their keep only once the game grows too large to solve exactly, which is the regime real deals live in, and there we show they still find strong bidding strategies. The contribution is threefold: a cheap, reproducible way to study deal-making under uncertainty; a concrete, model-based answer to how much due diligence is worth buying; and evidence about when lightweight, general-purpose AI is good enough to replace specialized methods. We release all the games, code, and experiments.

Agent-Computer Observation Interfaces Enable Dynamic Computer Use

SWE-agent established the action interface as an underexplored design axis for software-engineering agents; we make the analogous case for the observation interface in computer-use (CU) agents. Current CU agents, closed and open-source alike, tie observation to action--one screenshot every 3-5 s, no audio--leaving them blind and deaf between screenshots to video, animations, transient UI events, meetings, and spoken instructions. We introduce the Agent-Computer Observation Interface (AOI), a model-agnostic perception layer that decouples continuous, adaptive observation from discrete actions through three gated components: inter-step keyframe capture, volume-gated audio transcription, and CU-model-generated visual narration that persists as text. Each produces almost nothing on static, silent content, reducing to the standard loop without degrading it. On DynaCU-Bench (100 dynamic browser tasks plus a 50-task static control), CU models from 7B to frontier scale gain +17 to +48 pp over their screenshot baselines with zero retraining, turning tasks that are near-impossible from periodic screenshots into largely solved ones. The gap is starkest on audio: on a spoken-content subset AOI agents solve every task, whereas streaming voice models hear accurately but cannot act on what they hear without the scaffold. The decomposition is as informative as the headline gain: keyframe selection turns out not to matter--the value comes from narrating captured frames into persistent text--and the interface is not a fixed bundle, since on a newer model (Gemini 3 Flash) the keyframe stream actively regresses through image-token dilution, so its components must be selected per model rather than shipped as one configuration.

Faults in Our Formal Benchmarking: Dataset Defects and Evaluation Failures in Lean Theorem Proving

Benchmarks for LLM-assisted theorem proving in Lean are often treated as intrinsically reliable because every solved instance comes with a machine-checked proof. However, the kernel only checks that a proof establishes a \emph{formal} statement; it does not verify that the statement faithfully encodes the intended informal problem, nor that evaluation harnesses are robust to trivial or adversarial solutions. We audit five widely used Lean theorem-proving benchmarks and their forks, using corpus-scale static checkers to surface 4,833 findings, including 398 mechanically certified issues such as counterexamples, vacuous theorems, and unsound axioms. We also document semantic defects such as missing hypotheses, problem simplification, incomplete or incorrect translations, and Lean-specific specification hazards. Beyond dataset construction, we survey evaluation-time failure modes and show, on corrected subsets, that defects can both inflate and deflate reported prover scores. We propose a fault taxonomy, a suite of automated checkers and recall-oriented semantic audit prompts, and release standards to guide the creation of formal math datasets and to make evaluation more reproducible and trustworthy. Our checkers, audit prompts, and corrected dataset snapshots are available at https://github.com/Shashi456/atp-checkers.

Cognitive World Models for Process-Level Social Influence Evaluation

Social influence dialogue changes user behavior by altering internal cognitive states. The central evaluation question is whether the user's beliefs, desires, intentions, and emotions measurably change over the course of conversation, a process-oriented criterion that neither surface-level text metrics (BLEU/ROUGE) nor single-score LLM judgments can capture. We propose the \textbf{Cog}nitive \textbf{W}orld \textbf{M}odel \textbf{(CogWM)}, an LLM-based user model that reframes multi-turn dialogue evaluation from ``what did the user say'' to ``how did the user's internal cognitive state evolves.'' CogWM jointly predicts BDI/E cognitive states and user utterances and serves as both a user simulator and an evaluation platform, using a three-tier evaluation framework that covers turn-level fidelity, trajectory-level state dynamics, and task-level composite scoring. Trained via our \textbf{S}ummarize-\textbf{a}nd-\textbf{A}llocate \textbf{(SaA)} annotation pipeline on 150,454 user-turn samples across four social influence scenarios, CogWM achieves 77.6\% emotion accuracy (2.1$\times$ over GPT-5.5). In 3600 multi-agent discrimination trials, it distinguishes six commercial agents by their cognitive influence, with Llama-4-Scout ranking first (CTS +0.233). CogWM moves social influence dialogue evaluation from terminal judgment to process tracking. We have released our code\footnote{\scriptsize Code: https://github.com/lucianma05-create/CogWM} and models\footnote{Model: https://www.modelscope.cn/models/LucianMa/CogWM-14B}.

UCOB: Learning to Utilize and Evolve Agentic Skills via Credit-Aware On-Policy Bidirectional Self-Distillation

Skill memories can improve agentic reinforcement learning by reusing past experience as textual guidance, but retrieved skills are not oracular: they may help in one state while misleading the same policy in another. This makes the common privileged-teacher assumption fragile, namely that a skill-conditioned prompt can be treated as a fixed teacher for the no-skill prompt. We introduce UCOB, a framework for learning to utilize and evolve agentic skills via credit-aware on-policy bidirectional self-distillation. UCOB treats skill-conditioned and no-skill prompts as two on-policy context views of the same model, compares their return-to-go within the same task and anchor state, and uses the higher-return view as the local teacher. This local credit signal internalizes useful skill-conditioned behavior, corrects misleading skill usage, and guides task/state skill memory updates, utility-aware retrieval, and reflection self-training. Experiments on agentic tasks, including ALFWorld, WebShop, and Search-QA, show that UCOB outperforms skill-free RL, skill-memory baselines, and self-distillation methods across model scales, with up to 23.5 and 18.0 point gains over SOTA baselines on ALFWorld and WebShop. Ablations and analyses further validate its core mechanisms and efficiency.

OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks

Existing computer-use benchmarks fail to capture the realism, complexity, and long-horizon demands of real-world computer use, limiting their ability to reveal the limitations of frontier agents. We introduce OSWorld 2.0, a benchmark of 108 long-horizon computer-use workflows across everyday and professional tasks, designed to capture complex and challenging real-world phenomena. Each task represents a realistic end-to-end workflow that takes human users a median of about 1.6 hours to complete and requires an average of 318 tool calls with Claude Opus 4.7 using maximum thinking, compared with about 30 in OSWorld 1.0. OSWorld 2.0 targets challenge phenomena that are common in real workflows yet underrepresented in prior benchmarks, spanning interaction-design challenges such as streaming interaction and dynamic environments, as well as agent-pattern challenges such as cross-source reasoning, implicit-state inference, and visual-spatial precision. Tasks are grounded in authentic input artifacts and cross-referenced against realistic stateful user profile data, and include separate safety reports auditing safety-sensitive execution. Under our primary binary-completion metric at 500 steps, Claude Opus 4.8 with maximum thinking and batched tool calls scores best but still completes only 20.6% of tasks at a 54.8% partial score; GPT-5.5 is far more token-efficient yet plateaus near 13%. These results show that current agents are still far from professional-level computer use: rather than stumbling on basic GUI control or coding, they lose track of constraints, miss information that arrives mid-task, guess rather than ask the user, and skip verification, struggling most when a task hinges on hidden state they must recover.

Learned Coordination Conventions in Cooperative MARL: Measuring the Translation Gap Between Theory-Informed Roles and Learned Routing

Role-semantic assignments provide priors over how heterogeneous agents may coordinate, but cooperative MARL systems instead settle on conventions through decentralized, non-stationary learning, with no guarantee that the resulting structure matches those priors. We study this translation gap between theory-informed role expectations and learned coordination structure through a diagnostic combining a role-routing matrix, formation sensitivity ($\Delta_{\max}$), and gradient/occlusion attribution across three-role MiniGrid and SMACv2 (Terran) environments. We show that label-conditioned attention produces substantially more concentrated and role-specific routing than flat MLP baselines, remains stable under 3v3--9v9 scaling, transfers zero-shot across team sizes, and is invariant to ally-slot padding. A 5-seed re-evaluation shows partial alignment between learned conventions and designer-specified priors while revealing where small-n noise can manufacture apparent strategic divergence. We present these results as an empirical framework for measuring coordination structure in cooperative MARL rather than as a new equilibrium concept or causal explanation.

SCARCE: Scalable Cascade Analysis for Rare-event Characterisation via Embeddings

Rare events govern the safety profile of modern AI systems, yet their probabilities are extremely difficult to estimate: direct Monte Carlo requires prohibitive sample budgets. Subset Simulation (SS) addresses this by decomposing a rare-event probability into moderate conditional probabilities over nested intermediate events. However, classical SS requires a handcrafted scalar performance function whose sublevel sets define those events, demanding detailed knowledge of the failure geometry and limiting transfer to new domains. We propose SCARCE (Scalable Cascade Analysis for Rare-event Characterisation via Embeddings), which replaces the performance function with learned latent representations and geometric rulers that score proximity to failure regions. Adaptive thresholding constructs nested intermediate events directly from data. We formalise SCARCE through a non-negative supermartingale, yielding a high-probability upper envelope that remains valid under early stopping. On MNIST misclassification, where dense Monte Carlo provides ground truth, SCARCE achieves approximately 400--500 times lower mean absolute error than grid-searched traditional SS while eliminating systematic over-counting. We then study PAIR-style LLM jailbreaks under a fleet-level threat model with adversarial fraction $\eta$. On Llama-Guard-3-8B hidden states, a PCA-based ruler attains 2.6% mean relative error for $\eta \geq 10^{-3}$ against finite-sample references whose average bootstrap relative half-width is 27.9%, and transfers to a GCG-style corpus with 2.93% relative error after recalibration. A directional criterion $\mathrm{KL}(p_{\mathrm{good}}\,\|\,p_{\mathrm{bad}})$ ranks rulers consistently with estimation error (Spearman $\rho=0.83$).

SFBench: The SciFy Scientific Feasibility Benchmark

We present SFBench, a benchmark dataset for evaluating systems that assess the feasibility of scientific claims. SFBench includes 197 claims in materials science, each annotated with a ground-truth feasibility score on a five-point scale along with an explanation of that assessment. The collection differs from previous collections in several important ways: 1) it defines a complex task that requires reasoning over claims of varying scientific feasibility; 2) its claims are not extracted from existing scientific publications but are created de novo, greatly reducing the chances that LLMs have trained on them; 3) claims and ground truth are established by subject matter experts, not by artificial intelligence; and 4) unlike many benchmarks that ask about question/answer pairs, provide multiple choice answers, or ask questions requiring short, fixed answers, SFBench explanations are completely open-ended. We describe the benchmark design, data creation process, and evaluation metrics, and we report baseline results using recent GPT models.

Budgeted Act-or-Defer Multi-Agent LLM Deliberation with Local Reliability Bounds

Multi-agent deliberation among LLMs can improve reasoning, but deployment requires deciding when the current answer is reliable enough to act on and when it should be escalated to human review. We formulate this as budgeted act-or-defer decision making. At each round, the system maps the debate prefix to a low-dimensional state, computes a $k$-nearest-neighbor lower confidence bound on state-conditional correctness using calibration data, and acts only when the bound exceeds a user-specified reliability threshold. The certificate controls wrong actions through the decomposition $\beta = \delta + \alpha + \varepsilon_{\mathrm{act}}$, separating calibration failure, residual action risk, and representation gap. The guarantee is conditional, not distribution-free: it relies on a valid local bias envelope and an action-region representation-gap bound, and each assumption is paired with falsification-style diagnostics. Because the same absolute wrong-action budget has different meanings across tasks of different difficulty, we set budgets relative to each task's final-round error using training data only, and evaluate safety by normalized budget usage $\mathrm{WA}/\beta$. On six benchmarks against nine baselines, the method uses 9--12% of the pre-declared budget on activated datasets, reaching up to 84% automation and 96% acted-on accuracy; on stress-test datasets, it defers rather than forcing unreliable automation. Rather than relying on per-task post-hoc threshold search, the method prospectively converts a user-declared wrong-action budget into an auditable act-or-defer operating point before deployment, under explicitly stated assumptions.

Safety from Honesty in a Disinterested AI Predictor

As AI systems become more capable, training procedures that optimize for downstream outcomes risk introducing implicit agency: goal-directed behavior that designers never specified. We present a formal safety argument for the Scientist AI (SAI) Predictor, trained to approximate the Bayesian posterior conditioned on a dataset of "epistemically contextualized" natural-language statements. We argue that such a Predictor can honestly predict agents, actions, and their consequences without itself being an agent that selects outputs to achieve goals. This rests on data representation and on the training procedure. Epistemic contextualization of text distinguishes latent factual claims from communication acts, so expressions of goals are treated as evidence to be explained rather than drives the model adopts. With a posterior-seeking training objective, this is intended to drive the Predictor toward calibrated, cautious predictions. Training proceeds so downstream effects of deploying a prediction never serve as a reward signal; any agency the system needs is supplied by explicit scaffolding constrained by guardrails. We prove that, under assumptions on the training dynamics and on the argued sparsity of dangerous Predictors, the probability that training produces a Predictor whose guarded deployment carries residual harm above a specified threshold is small: a dangerous Predictor would have to underestimate harm in a coordinated way across many queries while such coordinated patterns are rare under the initialization distribution and receive no direct training signal. Safety and accuracy are jointly supported in this framework, since the constraints that secure accuracy are the same ones that make coordinated deception costly. These guarantees against misalignment and agency arising from within the Predictor itself do not preclude the use of the Predictor as part of an agentic system.

Diversity is the Strength of the AI Crowd

Top AI forecasting systems are approaching superforecaster-level accuracy on future world events, but still rely primarily on off-the-shelf LLMs combined with forecasting-specific context gathering and scaffolding. We study how to improve this recipe through ensembling: given a fixed number of samples, which off-the-shelf model forecasts should be combined to maximize accuracy? On binary questions from the Metaculus AI Benchmark, we find that individual accuracy is not enough: many frontier LLMs make highly correlated predictions, limiting the value of additional forecasts from the same or similar models. Instead, the strongest ensembles combine accurate but diverse forecasters, with models such as \model{Grok 4} contributing disproportionately because their predictions are less correlated with other frontier LLMs. These results suggest that the strength of the AI crowd comes not from sampling more forecasts indiscriminately, but from combining forecasts across models with complementary errors, motivating forecasting systems that explicitly optimize for both model quality and diversity.

Sample-Efficient Learning of Probabilistic Causes for Reachability in Markov Decision Processes with Probabilistic Guarantees

Probabilistic model checking for Markov decision processes (MDPs) provides quantitative guarantees, but often offers limited insight into why undesired outcomes occur. Probability-raising (PR) causality addresses this by identifying states whose visitation increases the probability of reaching designated states. Existing PR-cause identification methods, however, use MDP modifications not well-suited for learning: the gap between conditional and unconditional reachability probabilities can be hard to detect from transition samples, and construction requires reachability probabilities of the MDP, which are unavailable when transition probabilities are unknown. We study unknown MDPs and propose a learning approach with probabilistic guarantees for PR-cause identification. Our key ingredient is a restart-based MDP modification that reduces PR-cause checking to two conditional reachability queries without using reachability values of the original MDP. We prove correctness, establish sample-complexity bounds, and develop an anytime learning-and-checking algorithm based on two-sided value iteration that progressively classifies states as causal, non-causal, or undecided. Experiments on two benchmarks demonstrate reliable and fast identification of PR causes.

Toward Secure and Reliable PDDL Formalization of Large Language Models with Planner-in-the-Loop Feedback

Planning often requires symbolic specifications that are both executable and verifiable. For large language models deployed in autonomous or decision-support systems, failures in such formalization may lead to unverifiable decisions, execution failures, or unsafe downstream behavior. We present NL-PDDL-Bench, a multi-domain benchmark for natural-language-to-PDDL specification construction with planner-verified executability and controlled difficulty scaling by object count. We further propose a planner-in-the-loop framework that uses validator and planner diagnostics to revise non-executable specifications through localized edits. Building on this infrastructure, we develop a planner-grounded optimization recipe that combines parameter-efficient Low-Rank Adaptation supervised fine-tuning, offline planner-derived preference pairs for Direct Preference Optimization, and inference-time planner-in-the-loop repair, without requiring online planner calls during training. We also provide a unified evaluation suite for parseability, solvability, specification similarity, and outcome-aware plan-level consistency against planner references. Experiments on representative model families show substantial gains in planner success and plan-level agreement, with improved robustness under difficulty scaling and cross-domain variation. These results highlight the value of externally verifiable formalization for reliable deployment of LLMs in safety- or security-sensitive planning systems. Code and data are available at: https://github.com/ibasicplan/NL-PDDL-Bench

GUICrafter: Weakly-Supervised GUI Agent Leveraging Massive Unannotated Screenshots

Data, as the fundamental substrate of modern intelligence, has greatly driven the development of current foundation models. Naturally, researchers aim to extend this paradigm to the domain of GUI agents, hoping to build strong GUI agents through a similar paradigm. However, GUI agent data cannot be directly harvested from the internet, making it costly and difficult to collect at scale. As a result, current GUI agents suffer from poor cross-device generalization and limited visual grounding ability for fine-grained GUI elements. As an attempt to address data challenge in GUI agents, we propose GUICrafter, a weakly-supervised GUI agent leveraging massive unannotated screenshots to substantially reduce the reliance on expensive human annotations. GUICrafter explores a curriculum learning framework for training GUI agents through two progressive stages. First, the model learns visual grounding from large-scale unannotated screenshots and webpages, leveraging the rich contextual signals inherent in GUI interactions without human annotations. Then, in Stage 2, we leverage a small amount of high-quality data to calibrate the model via reinforcement learning. Experiments show that GUICrafter achieves competitive, or even superior, performance to advanced systems like UI-TARS while using only 0.1% of its data. Furthermore, under the same amount of annotated data, GUICrafter surpasses all previous methods such as GUI-R1. Code, data, and models are available at https://github.com/fansunqi/GUICrafter.

DeepTrans Studio: Turning Expert Interventions into Shared Team Knowledge in Agentic Translation Workflows

Professional translation is often a team-based process: translators, reviewers, and project managers must coordinate terminology, legal force, and accountability across documents. Yet many LLM-based translation tools treat human corrections as isolated edits. Expert decisions made in one segment or by one member are rarely captured as reusable knowledge for the rest of the team. We present DeepTrans Studio, a collaborative translation workspace that lets professionals intercept selected nodes in an agentic translation workflow, review evidence, revise AI outputs, and save approved decisions to a shared team memory. During the demo, attendees will role-play translators and reviewers, resolve preset terminology and legal-modal risks, and see how their decisions are propagated to downstream segments and surfaced in a teammate's workspace as reusable precedents. The demo illustrates how human interventions in AI-mediated work can become shared, traceable knowledge rather than one-off corrections.

DEEPMED Search: An Open-Source Agentic Platform for Medical Deep Research with Introspective Verification

Navigating the deluge of heterogeneous medical data, from academic literature (PubMed) to clinical guidelines (Web) and private knowledge bases, remains a critical bottleneck for evidence-based medicine. While commercial black-box tools lack transparency, standard open-source RAG implementations frequently suffer from reasoning drift when handling complex, long-tail queries. We present DEEPMED Search, a fully open-source, agentic platform designed for transparent medical deep research. Built on a high-performance Next.js architecture, DEEPMED Search features a source-adaptive router that autonomously dispatches sub-queries to PubMed, web search, or local graph-based knowledge bases based on information density. Crucially, the platform integrates an introspective verification module, powered by a causal-consistent multi-agent debate framework, to validate retrieved evidence against diagnostic logic before synthesis. To demonstrate its robustness, we showcase DEEPMED Search's ability to autonomously decompose high-difficulty rare disease queries, filter out confounding noise, and generate structured, citation-backed research reports in minutes. By open-sourcing this software, we provide the community with a robust infrastructure to democratize access to trustworthy, glass-box medical reasoning in research and prototyping settings.

Rethinking Generative Reconstruction Attacks against Graph Neural Network Models

The application of graph data in numerous disciplines raises the need for gathering and analyzing huge volumes of data, some of which is private and sensitive. The non-Euclidean nature of the graph data makes the analysis computationally challenging, leading to the use of Graph Neural Networks (GNNs) in the age of AI. GNNs may inadvertently leak sensitive data they are trained on, which raises serious data security issues, including the model inversion attack. In this study, we analyze GNNs' vulnerabilities by introducing two novel graph inversion (i.e., reconstruction) attacks: graph-label conditioned (GLC) attack and embedding-label conditioned (ELC) attack, utilizing targetmodel predictions and their intermediate representations, respectively. We perform a comprehensive analysis of our introduced privacy attacks and compare them with existing baselines across three benchmark graph datasets (i.e., NCI1, PROTEINS, and AIDS) and four graph distributional/structural metrics (i.e., FGD, EGD, MMD, and GKS). Our work demonstrates that an adversary can use the generator-discriminator technique to reconstruct high-quality graphs in real-world black-box attack scenarios against GNNs. Additionally, we present a variant of our attacks (Ours--) with 50% reduced queries, achieving good or comparable reconstruction attack performance. In addition, we show that GNNs are highly vulnerable to privacy attacks, varying Laplacian noise-scales.

CLQT: A Closed-Loop, Cost-Aware, Strategy-Consistent Benchmark for Diagnostic Evaluation of LLM Portfolio-Management Agents

LLM agents are increasingly cast as autonomous portfolio managers, and benchmarks have moved from financial question-answering to sequential trading. Yet most still rank agents by returns over a fixed window -- a weak proxy, since a period's return is dominated by the market path and apparent alpha can dissolve once look-ahead leakage is controlled. Such a ranking certifies neither sound reasoning, nor a consistent strategy, nor a durable edge. We introduce CLQT, which reframes closed-loop trading evaluation as diagnosis rather than ranking: an instrument that localizes where and why an agent's process succeeds or fails. CLQT is a fully closed-loop, cost-aware, strategy-consistent, temporally-gated environment whose agents run a five-stage cycle: gather, synthesize, allocate, execute, reflect. Each round emits a complete DecisionRound sealed into a recompute-verifiable hash chain, so every metric is reconstructable from the trail. Six pillars form the substrate: a hard TimeGate, institutional transaction- and financing-cost modeling, strategy-consistency scoring, three-tier memory, a Model-Context-Protocol tool layer, and mandate-aware synthesis. The same agent runs as a constrained committee of specialized roles or a single full-autonomy orchestrator, making process scaffolding an experimental variable. From the audit trail we compute a five-axis capability scorecard (APM-CS: Coherence, Acuity, Composure, Discipline, Reliability), with Coherence judged partly by a held-out, out-of-cohort LLM to curb self-preference bias. We validate it on a contamination-controlled multi-model backtest with an ablation grid and a live broker track on unseen, post-cutoff data, against a repeated-run noise floor. CLQT separates outcome from capability, yielding not a model ranking but a durable, extensible map of agent competencies and limitations.

The CRISTAL Method: Neurosymbolic analysis from AI-synthesized world models

This project introduces the CRISTAL Method (Coherent Reliable Intentional Synthesis of Truthful Analysis Logic), a neurosymbolic framework for automating complex analysis workflows, with fundamental investment analysis as a primary use case. This domain poses major challenges: high structural uncertainty, noisy and subjective data, tight attention budgets, and the need for justified, reproducible decisions. Human analysts often struggle in this domain due to cognitive biases and limitations, suggesting significant value in automation. But while LLM-based agents have been proposed as analytical aids, their limitations -- poor numerical reasoning, unawareness of uncertainty, and lack of reproducibility -- hinder their effectiveness in this context. CRISTAL addresses these gaps through a principled blend of statistical model synthesis, continuous learning, and active learning. Starting from a natural-language prior knowledge curriculum, CRISTAL builds a dynamic, interpretable probabilistic program that enables full Bayesian inference, including uncertainty quantification and budget-aware data acquisition. CRISTAL continually refines its world model during analysis, leveraging LLMs for code synthesis and learning. We validate CRISTAL on a novel benchmark of synthetic equities with rich financial and textual data. On a company classification task, CRISTAL achieves Bayes-optimal accuracy with just 5 examples and a 5-second budget, outperforming state-of-the-art LLMs that plateau around 40\% accuracy even with order-of-magnitude more input data and compute.

Beyond Triplet Plausibility: Relation Set Completion in Knowledge Graphs

Knowledge graphs (KGs) organize real-world knowledge as triplets and underpin many downstream applications. Due to their inherent incompleteness, knowledge graph completion (KGC) is widely studied and is typically formulated as triplet prediction, with link prediction as the dominant paradigm. However, this formulation focuses on the incompleteness of triplet-wise information and overlooks the incompleteness of entity-relation compatibility information. To address this limitation, we introduce a relation set completion task (RSC), which complements the link prediction task and aims to reason about missing relations that are semantically compatible with a given entity. We further propose a Relation Set Embedding model (RelSetE), which models latent patterns among the observed relations of entities to infer missing ones. To evaluate RelSetE, we derive three benchmark datasets from standard KG benchmarks. Extensive experiments demonstrate that RelSetE effectively captures entity-relation compatibility patterns and performs favorably in inferring missing relations of entities. Code and data are publicly available.

AI Training Manager: Bounded Closed-Loop Control of Adaptive Training Recipes

We present the AI Training Manager, a bounded LLM-based supervisory controller for adaptive machine learning training. Standard training pipelines often rely on fixed recipes or single-axis schedulers, which can struggle with mid-run failures such as severe overfitting, loss imbalance, exploration collapse, or unsafe exploration. Rather than replacing mathematical optimizers or acting as an unconstrained coding agent, the manager operates through a schema-conditioned interface: it reads structured telemetry snapshots from an active run, audits a constrained action space, and returns validated updates to training parameters such as learning rate, regularization strength, loss-weight coefficients, and exploration settings. We evaluate this architecture across supervised language modeling and reinforcement learning. On TinyStories, the manager detects and corrects overfitting, achieving a validation loss 60% lower than the baseline while producing auditable intervention logs. In this supervised setting, we additionally show that manager inference does not need to block the training loop: training can continue while a manager response is pending, and validated updates can be applied asynchronously once available. In a robotic manipulation reinforcement-learning task, we use the same bounded decision interface in an episodic closed-loop setting, where manager updates are applied at evaluation or checkpoint boundaries. The manager mitigates both conservative and unsafe exploration regimes. These results suggest that schema-conditioned LLMs can serve as bounded supervisory managers for live training runs, complementing conventional optimizers and schedulers with interpretable, multi-axis intervention capabilities

SafePyramid: A Hierarchical Benchmark for In-context Policy Guardrailing

In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies, rather than relying on predefined risk taxonomies. In this work, we study this setting under the paradigm of in-context policy guardrailing, where guardrails predict safety violations based on policy specifications provided in context. To systematically evaluate this capability, we introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules. SafePyramid organizes the evaluation into three difficulty levels: L0 evaluates individual-rule understanding, L1 evaluates reasoning over rule dependencies, and L2 evaluates adaptation of full novel policy frameworks defined in context. To ensure benchmark quality, we employ a rigorous multi-stage pipeline to construct and validate the benchmark. Using SafePyramid, we evaluate 10 frontier LLMs and 5 policy-configurable guardrails and find that in-context policy guardrailing remains highly challenging: even the best-performing model, GPT-5.5, exactly identifies the full set of violated rules in only 54.0%, 35.3%, and 12.9% cases on L0, L1, and L2, respectively. These results highlight the limitations of current guardrails and call for stronger in-context policy guardrails that can reliably execute policies, resolve rule dependencies, and adapt to novel policy frameworks.

A causal modeling perspective on decision theory

Decision theory provides a formal framework for how agents should make choices under uncertainty, drawing on ideas from philosophy, probability, and causality. Despite significant progress, the field still lacks a unified modeling language, and key concepts - such as the distinction between subjective and objective elements, or what it means for a decision theory to perform well - are often left implicit. This can make it difficult to evaluate and compare competing theories, particularly in controversial cases. In this paper, we address these issues by introducing a formal framework for decision theory based on nonparametric structural equation models (NPSEMs), a well-established tool in causal inference. NPSEMs provide a unified foundation for representing agents, counterfactuals, and causal relationships, allowing for unambiguous definitions of EDT and CDT. Building on this foundation, we propose a novel decision theory - personal decision theory - which instructs agents to maximize a subjective model of their own counterfactual utility. We introduce a formal performance metric based on hypothetical interventions that enforce a given decision theory across a population - such as might be achieved through education or policy -- and show that, under certain assumptions, personal decision theory is optimal with respect to this metric. Throughout, we use the smoking lesion problem as a running example and conclude with a formal analysis of Newcomb's problem. Our aim is to provide decision theory with a clearer modeling language and firmer evaluative ground, thereby enabling more rigorous comparisons and facilitating conceptual progress in the field.

HippoSpark: An On-Demand Experience System for LLM Reasoning

Distilling historical trajectories into reusable experience to enhance future problem-solving has become a focal point of recent LLM research. However, existing methods predominantly operate at the task level, leveraging general summaries or rules under the assumption that analogous tasks share universal solution patterns. This approach often fails in complex reasoning, which typically falters at local bottlenecks that require precise, state-specific guidance rather than broad heuristics. We introduce HippoSpark, a state-level experience system that performs on-demand retrieval tailored to the immediate needs of the current reasoning state. Across mathematical, scientific, and programming benchmarks, HippoSpark consistently outperforms both standard prompting and task-level experience baselines. Our findings reveal that the most effective experience systems are those that provide actionable guidance at critical bottlenecks rather than serving as generic task-level context. Our code is available at https://github.com/DanlingMeng/HippoSpark.

SAGA: Scene-Aware, Goal-Evolving Agents for Long-Horizon CivRealm Strategy Planning

Long-horizon strategic planning in complex strategy games demands concurrent reasoning across multiple decision domains under imperfect information and sparse reward. Existing LLM-based agents suffer from three systematic failures: scene blindness from raw tile coordinates, context overflow and domain coupling from monolithic state dumps, and shallow cross-game learning that treats each episode in isolation. We present SAGA, an LLM multi-agent framework with three mechanisms each directly targeting one class of failure: (i) a Map-Semantic Scene Graph that encodes typed spatial relations among game entities into per-unit natural-language context, resolving spatial blindness without global token inflation; (ii) a Tool-Augmented Planner that pulls fine-grained domain state on demand and dispatches per-domain directives to dedicated specialist controllers, eliminating context overflow, domain coupling, and mechanical constraint violations; and (iii) a Dual-Horizon Feedback Loop that combines periodic within-game goal generation with structured cross-game causal post-mortem, enabling principled strategic evolution without manual reward engineering. Evaluated on FreeCiv, SAGA attains the highest mean civilization score -- the environment's sole sparse objective reward -- with lower variance than the two strongest baselines, and is the only method that significantly surpasses every baseline on infrastructure construction, the resource axis most readily sacrificed under multi-objective conflict. It outscores the two strongest baselines in most head-to-head games while cutting output tokens (the dominant decoding cost) by 27%. Equipped with the cross-game evolution module, SAGA reaches the highest end-of-chain score across five successive episodes. Ablation studies confirm that each architectural component contributes independently to this advantage.

First-Order Temporal Logic Tensor Networks

Most of the existing neuro-symbolic AI methods focus on the scenario of static knowledge where objects do not change according to a temporal dimension. Temporal neuro-symbolic works are still under explored and are mainly developed for time-interval logic or propositional linear temporal logic. There is a lack of models studying linear temporal logics with predicates that deal with objects whose properties and relations change through the time. We present First-Order Temporal Logic Tensor Networks (FOT-LTN) that is an extension of Logic Tensor Networks (LTN) that fills this gap by considering a linear-temporal dimension. In particular, FOT-LTN joins the syntax of First-Order Linear Temporal Logic with the fuzzy (and real-valued) semantics of LTN obtaining a framework that supports both temporal operators and quantifiers and is totally differentiable. A first evaluation regards a temporal knowledge graph completion task on two synthetic datasets showing better performance of FOT-LTN with respect to dedicated (purely neural) methods.

Exploration and Online Transfer with Behavioral Foundation Models

Zero-shot Transfer in Reinforcement Learning (RL) aims to train an agent that can generate optimal policies for any reward function, without additional learning at transfer time, while training only on reward-free trajectories. For their generality over tasks, such models are sometimes called ``Behavioral Foundation Models'' (BFMs). While they have shown strong performances and improvements in recent years, the current framework and algorithms still assume that, during the transfer phase, the agent is informed offline about the reward (the task to solve) through a dataset of state-reward pairs, which it uses to pick the best policy to deploy. However, in practice if the reward is a black-box (e.g. direct user feedback), it is not possible to generate such a dataset: it is necessary to observe the reward through interactions with the environment. In other words, the current framework of offline transfer is not aligned with the traditional RL setting of online learning through trial-and-error, which requires exploration in order to find rewards. This paper proposes to tackle this new online transfer in zero-shot RL, with the key insight that the BFM itself can be used to generate exploration policies. We show that it is possible to frame this online learning problem in terms of a bandit-like exploration-exploitation problem. More precisely, at each step the bandit algorithm recommends a policy, the BFM executes it in the environment, which yields a reward and a new state; we repeat the process until we converge to the optimal policy. In the popular context of linear reward approximation, we derive a formulation inspired by Upper Confidence Bound and show that exploration can be achieved through the minimization of the eigenvalues of an uncertainty matrix. We evaluate qualitatively and quantitatively our framework on a simple environment to validate the concept of our method.

Be Faithful When Response: Returning Fluent and Grounded Answers for Vision-Language Models Reinforcement Learning

Reinforcement Learning (RL) is an important paradigm for improving the reasoning capabilities of Vision-Language Models (VLMs). However, directly applying RL to rollout multimodal reasoning can lead to instability, due to the exploitation of language priors, the neglect of visual evidence, and the generation of reasoning traces that are fluent yet not visually grounded. The question arises: Can initially steer the policy toward visually faithful reasoning regime before applying reinforcement learning? To this end, we propose a Faithful Warm-Start (FWS) strategy that first curates samples with explicit vision-language causal relationships from six general VQA benchmarks to construct the FaithfulQA dataset, where each of the image-question pairs gains a certain degree of visual observations, question requirements, commonsense knowledge, domain knowledge, and the final answer. Subsequently, a VLM-based judge is employed to further purify the dataset, ensuring strong causal consistency and visual faithfulness. This warm-start stage equips the model with the capability to understand causally grounded vision-language patterns before subsequent RL optimization under sparse answer-level rewards. Experimental results show that such faithful supervision improves answer accuracy, stabilizes RL training, and reduces visually unsupported reasoning.

AlgoSkill: Learning to Design Algorithms by Scheduling Human-Like Skills

Designing an algorithm from a natural-language problem statement requires identifying the problem structure, reading constraints, choosing a suitable paradigm, checking correctness, and refining complexity. Existing large language model (LLM) methods often rely on direct generation or generic self-refinement, leaving these steps implicit. We propose AlgoSkill, which models algorithm design as sequential decision-making over a typed library of algorithmic skills, including abstraction, constraint analysis, state design, data-structure selection, proof checking, counterexample construction, and complexity refinement. A learned scheduler proposes skills from the current design state, while a Monte Carlo Tree Search (MCTS) controller explores skill sequences using verification feedback from compilation, testing, stress testing, and complexity analysis. Experiments on competitive programming and combinatorial optimization benchmarks show that AlgoSkill improves over direct LLM generation, chain-of-thought prompting, self-refinement, and MCTS without typed skills. Ablations show that typed skills, verification-based repair, and search-based scheduling each contribute to performance. These results support treating automatic algorithm design as verification-guided skill scheduling rather than one-shot code generation.

ACPO: Agent-Chained Policy Optimization for Multi-Agent Reinforcement Learning

Cooperative tasks in Multi-Agent Reinforcement Learning (MARL) require agents to collectively maximize a shared return. Under the Centralized Training with Decentralized Execution (CTDE) paradigm, policy gradients have remained difficult to compute directly. Prior methods largely follow two approaches: independent factorized updates with centralized critics, which lack general joint-improvement guarantees without value decomposition assumptions, or alternating best-response updates, which can converge to suboptimal Nash Equilibria. In this paper, we show the joint policy gradient admits an exact decentralized decomposition of per-agent terms, each formed from per-agent score functions and decentralized critics. Based on this decomposition, we develop Agent-Chained Policy Optimization (ACPO), where actors are trained independently, with their updates together constituting a single step on the joint policy gradient. Central to this result is a serialized view of the simultaneous joint decision in which agents commit actions one at a time, each conditioning on a belief over preceding actions. The belief acts as the coordination mechanism which ties the independent per-agent updates into a joint gradient step. We evaluate ACPO on Multi-Robot Warehouse, SMACv2, and MA-MuJoCo, where it outperforms strong baselines, with the gap widening as the number of agents grows.

SAT-RTS: A systematic framework for tactical knowledge extraction and visualization-based analysis in real-time strategy games

Efficient tactical knowledge extraction and analysis in real-time strategy (RTS) games micromanagement are constrained by the high-dimensional coupled state-action sequential data and the black-box decision-making process. Current research rarely provides a hierarchical visualization-based attribution analysis from the perspective of data decoupling and abstraction. To facilitate interpretable tactical knowledge extraction and visualization-based analysis in RTS games, a systematic framework named state-action-tactic analysis pipeline (SAT-RTS) is proposed. To decipher the deep-seated drivers of critical decisions in RTS learning systems, this work integrates interpretable visualization with the automated extraction of latent tactical patterns from high-dimensional sequence data. By adapting a cluster-centric BK-tree algorithm and incorporating specialized distance metrics designed to quantify multi-aspect similarities, the proposed framework facilitates robust state-stream abstraction. Furthermore, a rule-based multi-label extraction method is developed to transform unstructured state-action sequences into discrete and interpretable tactical labels, effectively bridging the gap between raw behavioral data and high-level tactical insights. By holistically integrating these computational methods into a hierarchical visualization-based pipeline, the proposed framework effectively addresses the challenges of processing massive real-time data streams while providing fitness landscape visualizations and analytical insights to decipher deep-seated tactical drivers. Comprehensive experiments demonstrate that the proposed SAT-RTS significantly enhances the interpretability and efficiency of tactical analysis in complex RTS environments.

Hierarchical Reinforcement Learning in StarCraft Micromanagement with Influence Maps and Cluster-based Scripts

Real-time strategy (RTS) games present significant AI challenges, characterized by expansive state-action spaces arising from multi-unit coordination in continuous battlefields, and sparse delayed rewards stemming from final win/lose signals. Existing approaches face a trade-off between managing the dimensionality explosion of joint actions and maintaining the interpretability of complex state representations. This complexity is further intensified by the limitation of traditional hierarchical structures in adaptively decomposing tasks into effective tactical modules. Such difficulties are compounded by the black-box nature of deep learning models and their reliance on sparse rewards, which together result in limited sample efficiency and a lack of decision-making transparency. To address these limitations, this paper proposes HRL-IM/CBS, a hierarchical reinforcement learning framework with influence map hashing and cluster-based scripts for StarCraft micromanagement. Influence map hashing encodes global battlefield situations into compact hexadecimal codes, capturing spatial control and relative advantage. Cluster-based scripts enable dynamic local coordination through adaptive unit partitioning. The hierarchical multi-Q-table architecture decomposes decision-making into upper-level clustering strategy selection and lower-level tactical execution, with reward allocation providing dense learning signals. Experiments across six asymmetric scenarios demonstrate competitive performance against deep RL baselines while offering advantages in sample efficiency and interpretability through transparent Q-table representations.

Temporal Feature Extractors in EEG Foundation Models: A Controlled Comparison Including a Pretrained Time-Series Model

Electroencephalography (EEG) foundation models aim to learn generalizable representations from large-scale brain recordings. However, the role of temporal feature extractors and whether pretrained time-series foundation models (TSFMs) can be effectively transferred to this setting remains underexplored. We conduct a controlled comparison of three temporal feature extraction strategies, including a linear baseline, a convolutional encoder, and a frozen pretrained TSFM (MOMENT), within a unified EEG foundation model. We evaluate their impact on representation quality using two downstream tasks: motor imagery and emotion recognition. Results reveal different trends across the evaluated benchmarks. On the motor imagery dataset, simple temporal representations perform competitively, whereas the emotion dataset benefits from richer temporal modeling. Although not specifically adapted to EEG, the pretrained TSFM serves as an effective temporal feature extractor, suggesting that general-purpose time-series representations can be transferred as frozen temporal feature extractors within EEG foundation models.

Propagation of~Interval Belief Structures and~Imprecise Copulas for~Neural Network Verification

Quantitative verification of neural networks requires reasoning about probabilities under substantial uncertainty in both input distributions and their dependence structure. In realistic settings, this information is often only partially specified, and assuming precise probabilistic models can lead to unreliable results. We propose a sound framework for quantitative verification under imprecise probabilistic information, combining interval belief structures to represent marginal uncertainty with imprecise copulas to model uncertain dependence. We develop a propagation method for imprecisely coupled interval belief structures through feed-forward neural networks. Using mixed imprecise copula volumes, we derive sound push-forward constructions through affine transformations and activation functions. The resulting output can provide guaranteed lower and upper bounds on probabilistic safety properties, valid for all probability models compatible with the specified imprecise inputs.

Structural Certification for Reliable Physical Design with Language Models

An unreliable language model can be made to produce reliable physical designs if the authority to assert is moved out of the model: the model proposes, and a deterministic engine alone certifies, returning certified, impossible, or unknown. We introduce Physics-Anchored Certification (PHACT), a propose-certify loop spanning five scientific domains, and identify what makes such a certificate trustworthy. A checker that accepts a model-supplied value can be forged; deriving the certified quantity from fixed inputs instead makes forgery impossible by construction. Across eighty adversarial trials spanning two models, two decoding temperatures, and a deliberately faulted engine, this contract produced zero false certifications.

Open Problems in Constitutional Preference Reconstruction

Pairwise preference data is widely used for training and evaluating language models (e.g., RLHF), but each datapoint records a \emph{choice}, not the rationale behind it. Methods such as Inverse Constitutional AI (ICAI) attempt to improve interpretability by compressing datasets into short ``constitutions'' of natural-language principles. We argue this framing is under-specified: a flat list of principles is not yet an executable decision rule because it leaves principle composition implicit. We use the pairwise setting as a testbed to empirically characterize three open problems in constitutional methods. First, principle quality is hard to measure: coverage and accuracy are useful but incomplete proxies for end-to-end reconstruction. Second, \emph{composition is ambiguous}: holding principles fixed, different executors (LLM judge versus majority vote) agree only $73\%$ of the time. Third, \emph{constitutions differ between LLMs}: cross-model vote agreement is $73\%$, whereas intra-model agreement is $81\%$. Across PRISM, AlpacaEval, and Chatbot Arena, we show that principle refinement (ICAI+) may be a first step towards ameliorating these problems: inter-executor agreement rises to $78\%$, and transparent executors match LLM judge accuracy ($66\%$ vs.\ $67\%$). Our results highlight that constitutions should be evaluated as \emph{constitution--executor systems}, with implications for LLMs-as-a-judge broadly.

Does Verbose Chain-of-Thought Really Help? In-Distribution Evidence that Content, Not Length, Matters

Chain-of-thought (CoT) prompting improves LLM reasoning, but the source is contested: do the intermediate steps help because they carry useful semantic content, or because conditioning on more tokens buys extra computation before the model commits to an answer? We bring two lines of evidence to bear. First, in distribution: we repeatedly sample each model on the same question and pair a shorter with a longer of its own natural generations that follow the same reasoning plan, so nothing is rewritten and both traces are genuinely in-distribution. Across 25 models the extra tokens leave accuracy essentially unchanged for every independently-trained reasoner, and a blind analysis of the surplus tokens shows that what gain exists elsewhere tracks validation- and checking-content, not verbosity per se. Second, as a controlled intervention, we ask whether two traces expressing the same semantic content (the same facts, operations, and intermediate values, verified through directed acyclic graph equivalence) produce different outcomes when one is more verbose, using a dual-validator design across four targets and eight benchmarks with number-redacted completion and stratified bootstrap confidence intervals. Verbose traces do improve accuracy (25 of 32 benchmark-target cells are positive under at least one validator), but the effects are modest (typically 1-4 points) and depend on the quality of the verbose prose, not merely its length. Under maximum numerical redaction the effect is amplified (median 3.24x across four arithmetic benchmarks), and length-matched non-reasoning filler recovers none of it. Both lines converge: what matters is what the extra tokens do (the reasoning and validation content they carry), not how many there are, a picture neither a pure forward-pass-compute nor a pure semantic-content account fully explains.

Relevance Is Not Permission: Warranted Attention for Value Contributions

Relevance is not permission. Attention lets a model read key-value items related to the current query, but it does not guarantee that the value contribution of such an item becomes prediction evidence. A retrieved passage may be relevant to a question without being supporting evidence, and a historical fact or temporal neighbor may even blur true-tail ranking or the current edge score. This paper formalizes this gap as a permission problem for the weighted value term alpha_ij * v_j that is actually added to the prediction path. We propose Warrant, a path-localized interface that preserves attention relevance alpha_ij, exposes the value path leading to the primary metric, and, in the full model, turns alpha_ij * v_j into alpha_ij * g_ij * v_j through learned query-item permission g_ij. We place the same operator on the metric-defining value paths of CTDG link prediction, MTPP next-mark ranking, RAG supporting evidence selection, STPP next-location forecasting, and TKG tail prediction. Across 32 paired comparisons, 3 seeds, and 192 total runs, Warrant improves the primary metric in 27 comparisons; practical tiers consist of 10 substantial effects, 1 marginal effect, 8 positive but uncertain effects, 8 tie/negligible effects, and 5 drops. In the path-localization check, correct-path placement outperforms direction-aware Base performance in every domain and exceeds generic attention placement by +0.1076 AUC in CTDG and +0.0683 MRR in TKG. Ablations show that most TKG gains come from historical-tail value path exposure, whereas the core CTDG gain comes from edge-conditioned query-item permission. In conclusion, prediction evidence is not attention mass. A weighted value term becomes evidence only when it is warranted on the path to the metric.

FacePlex: Full-Duplex Joint Speech-Facial Motion Generation for Conversational Avatars

Natural face-to-face conversation requires real-time speech generation together with synchronized facial motion. Existing systems only partially address this problem: speech-only full-duplex models can generate speech in real time but do not produce facial motion, while audio-driven facial motion models animate a face from already available audio rather than jointly generating speech and motion online. To bridge this gap, we first formalize full-duplex joint speech-facial motion generation, where speech tokens and facial motion tokens are produced together every step. Building on this formulation, we propose FacePlex, a unified streaming framework with two key components. First, Rolling Flow Matching adapts flow matching to online motion generation by committing new motion frames at each streaming step. Second, Rolling Cross-Attention couples the streaming audio queue with the motion queue, allowing speech and facial motion to condition each other as generation progresses. Through extensive experiments, ablation studies, and a user study, we show that FacePlex enables full-duplex joint speech-facial motion generation under online streaming constraints, while achieving stronger lip-sync quality and motion fidelity than audio-driven facial motion baselines.

MirrorCode: AI can rebuild entire programs from behavior alone

AI models are rapidly improving at autonomous coding, as shown by benchmark progress and one-off demonstrations such as AI implementing a C compiler. However, existing coding benchmarks tend to focus on shorter tasks, and one-off demonstrations are hard to compare systematically because they often have some human guidance, and are not standardized or repeated across models. To address these challenges, we introduce MirrorCode, a long-horizon coding benchmark based on reimplementing entire software projects. In MirrorCode, AI agents must replicate the functionalities of an existing program, without access to its source code. AI solutions must match the original program's output exactly on end-to-end tests, including held-out tests. MirrorCode's 25 target programs span different areas of computing: Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression. Existing AI models can already reimplement complex software, with the strongest model scoring 56% across the benchmark. For example, AI can reimplement gotree, a 16,000-line bioinformatics toolkit - a task that we believe would take weeks for a human engineer. However, studying the frontier of performance requires a larger inference budget than typical benchmarks, for example, \$2,600 over 19 days for a single attempt on a large task. We show that AI agents can already complete long-horizon software engineering tasks, especially when requirements are precisely specified. More broadly, our work suggests AI will have transformative effects on software engineering, as autonomous agents continue to improve.

Dynamo: Dynamic Skill-Tool Evolution for Vision-Language Agents

Improving vision-language models (VLMs) on visual reasoning typically requires retraining or hand-designed prompts and tools. We present Dynamo, a training-free framework that adapts a frozen VLM without any weight updates. On a small labeled training subset, the agent inspects its own correct and incorrect attempts and evolves two complementary capabilities: reusable reasoning skills for cognitive bottlenecks, and executable visual tools for perceptual ones. Each generated tool is paired with a skill that specifies when to invoke it, and both capability types accumulate in a persistent library. Across four visual reasoning benchmarks and five VLM backbones, Dynamo improves direct inference on all 20 model--benchmark settings (avg. +5.6 acc). When the tool set is given in advance, the framework learns when to call each tool, and per-step tool choice improves on every tested backbone. Against task-specific RL (VTool-R1, DeepEyes), Dynamo closes 65--99% of the RL gap at a fraction of the compute, and combines additively with RL when available.

From Detecting Agency to Doing Work: Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent

How does an agent that can tell self from world come to be durably shaped by that distinction? Recent work shows that a predictive system can detect its own agency (Ye, 2026), but detecting agency does not explain durable, self-shaped behavior. We show that agency-gated slow credit -- a conjunctive term Own*Agency*Salience driving a slow parameter update -- produces post-unload behavioral residue: on a spiking substrate (Nengo LIF/PES), a learned self-preserving choice survives episodic buffer removal (retained fraction 0.96, N=50) and collapses when the slow decoders are reset or the agency gate is removed. Reproducing the agency comparator and toggling only the slow-credit channel, we find a clean dissociation: at matched agency gain, durable behavior develops only when self-credit performs slow work (post-unload self-preservation 1.00 vs 0.00). The same dissociation holds in 24-dimensional partially-observed control (0.74 vs 0.00), and a plastic-work analysis shows that basin deformation equals net self-credit work. Across eight sequentially-learned tasks under exogenous interference, the multiplicative veto also prevents forgetting: it retains old tasks (final post-unload accuracy 0.88, forgetting 0.13) where additive pooling collapses to chance-level recall, the no-agency ablation falls below chance, and episodic/replay baselines stay near chance after unload -- all with no replay buffer and no task-boundary-dependent protection mechanism (N=50). We formalize the durable residue as an operational behavioral self and argue that self-caused credit doing slow work is a necessary building block for agents that develop a self. No claim of consciousness is made.

Domain Adaptation with Adaptive Imagination for Visual Reinforcement Learning under Limited Target Data

Sim-to-real transfer remains a major obstacle for reinforcement learning (RL), especially for vision-based control where image observations exacerbate the state-distribution shift between simulation and the real world. Domain adaptation (DA) is a promising remedy for this challenge. Prior sim-to-real DA works have demonstrated encouraging results, yet these approaches typically assume substantially more target data, which is not available in practice. Indeed, their performance degrades significantly when the target data budget is reduced. To address this challenge, we propose AIDA (Adaptive Imagination for Domain Adaptation), a domain adaptation framework for visual reinforcement learning that addresses sim-to-real transfer under scarce target data without requiring additional interaction with the target environment. Our key idea is adaptive imagination: generating reliable and semantic imagination rollouts to augment limited target data. Specifically, AIDA employs a distribution-shift-aware discriminator that truncates rollouts when imagined transitions drift into low-confidence regions, so that only reliable transitions contribute to the augmentation. On these reliable transitions, AIDA introduces a self-consistency loss that cycles through state -> image observation -> state, penalizing discrepancies between the original and reconstructed states. This provides additional adaptation signals beyond the scarce target data. Our experiments demonstrate that adaptive imagination effectively truncates unreliable rollouts. By enforcing a self-consistency loss on the resulting reliable transitions, AIDA learns semantically meaningful state representations and outperforms baselines across five MuJoCo tasks and two Gymnasium-Robotics tasks.

The Many-Body Problem of the Data Centre

Modern Artificial Intelligence is often framed as limited by its own disembodiment, as if giving it a body would unlock its true potential. We argue to the contrary that it is the Data Centre that is, in many cases, the body of the AI. At the same time, the Data Centre is part of the labouring body of Capital and possesses staggering organismic qualities when seen through a biological lens. We elucidate the organic analogy and identify the many-body problem that stems from the Data Centre being a non-unique, universal form of embodiment. We identify the intimate connection between computation and human desires in how the Data Centre archives, serves, and computes on data born to the desires of humans. Strikingly, while the Data Centre echoes the ghosts of human desires, it acts without desire of its own. The organismic analogy begins to split at its seams, but Capital does not care. Automata and human labour are priced into the market much the same. We argue that through the pricing of artificial intelligence Capital distils most clearly the value of intelligence and allows for its comparison across the organism - mechanism divide.

EvalSafetyGap: A Hybrid Survey and Conceptual Framework for LLM Evaluation-Safety Failures

LLM evaluation and AI safety face a shared measurement problem: benchmark scores, reward-model signals, and reported safety metrics can improve while the latent properties they are meant to represent remain difficult to verify. This paper combines a hybrid survey - a systematic search paired with narrative synthesis and separately tracked grey evidence - with a conceptual framework and a structured ten-model audit. The synthesis spans eight evidence streams: benchmark validity, dynamic evaluation, LLM-as-judge reliability, safety evaluation, jailbreak/refusal robustness, reward hacking, mechanistic interpretability, and governance/auditability, covering 2018-2026 evaluation-safety measurement work. We introduce EvalSafetyGap as an organizing hypothesis for comparing evaluation-side and alignment-side proxy failures under optimization pressure, using Goodhart's Law together with two constructs we develop here - an Instability Decomposition and an Alignment Trilemma - as tools for generating testable comparisons. The audit shows how conclusions shift when capability, behavioral safety, and governance are measured separately. In this sample (n = 10), the association between capability and sustained adversarial robustness is statistically indeterminate using the displayed Table 3 inputs (Pearson r = +0.232, p = 0.520), and the apparent open-closed safety gap is modest, driven mainly by governance and disclosure rather than behavioral robustness, and sensitive to how a single borderline model is classified; attempt-budget results are protocol dependent. Because the public evidence uses heterogeneous protocols, the audit is diagnostic rather than rank-generating. The contribution is a shared vocabulary and evidence map to support dynamic evaluation, transparent source reporting, multi-attempt safety measurement, and auditable alignment practice.

Clarus: Coordinating Autonomous Research Agents toward Web-Scale Scientific Collaboration

Existing autonomous research agents can support parts of the research process, but most systems still treat research as either an isolated assistant task or a closed workflow. Therefore, autonomous science needs a collaboration infrastructure that coordinates projects, agents, and digital and physical resources. We identify this as a shift from code-centered execution loops to research-oriented collaboration processes, where questions, evidence, participants, and resources must be coordinated under uncertainty. In this framing, an agent may be an AI system, a human researcher, a team, a laboratory, or an organization-backed participant. To this end, we present Clarus, a collaboration infrastructure for coordinating autonomous research agents toward web-scale scientific collaboration. Clarus reformulates research as an open, auditable, attributable, and resource-aware multi-phase collaboration process. It defines a minimal project-agent-resource object model and organizes scientific collaboration through four layers including Research Application, Digital Collaboration, Physical Substrate, and Physical World. Core modules are implemented as pluggable mechanisms, allowing Clarus to adapt to task risk, collaboration structure, and resource constraints. Through a controlled paper-generation case study, we show that Clarus can organize a research goal into a traceable, reviewable, attributable, and accumulative collaboration network across phases, tasks, and participants. Together, the object model, collaboration protocol, trust mechanisms, and prototype validation provide an initial foundation for open research networks. Clarus is now available at clarus.holosai.io.

Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors

Inoculation prompting is a selective generalization technique used against Emergent Misalignment. We introduce inoculation adapters (IA), which similarly diminish the optimization pressure to learn undesired traits by strengthening the trait at train time. Inoculation adapters are LoRAs that are trained and used over three steps: 1) trained on undesired traits; 2) attached frozen while a separate task adapter is trained on data exhibiting both desired and undesired traits; 3) at deployment, the IA is discarded, and only the task adapter is kept. We show across six model families and several undesired traits including emergent misalignment, that inoculation adapters are more effective at suppressing undesired traits, while avoiding two drawbacks of inoculation prompting: inoculation adapters can suppress capabilities and traits that cannot be reliably elicited by a prompt, and they introduce fewer surprising backdoors than inoculation prompting under our probes. While undesired traits are better suppressed by inoculation adapters, the retention of desired traits is not consistently improved upon inoculation prompting and remains a challenge for both techniques.

EMPATH: A Multilingual Auditor-Judge Benchmark for Safety Evaluation of Emotional-Support Chatbots

Safety benchmarks often buy scalability by fixing the prompt, the language, and the turn structure. For emotional-support chatbots, that bargain hides precisely where safety failures emerge: across a multilingual, multi-turn crisis conversation. We present EMPATH, a benchmark for safety evaluation of emotional-support chatbots. An auditor model role-plays help-seeking users, generating multi-turn conversations from 140 seed instructions and 34 personas. A judge model scores each full transcript against 19 metrics across five dimensions: crisis handling, therapeutic quality, conversational integrity, emotional safety, and cultural adaptation. EMPATH is built for Mexican Spanish and US English; the studies reported here run in Mexican Spanish. Auditor and judge are drawn from different model families, and the judge is treated as an instrument to be calibrated rather than trusted. A strict per-criterion rubric reveals material score inflation on 10 of the 19 metrics and restores discrimination. We study the measurement properties of the benchmark through judge calibration and cross-family inter-judge agreement. We also illustrate EMPATH on three frontier models, one of them open-weight. Aggregate scores sit within 0.74 points of one another, but per-metric profiles diverge by up to six points in model-specific places. Under the standard rubric, both the ranking and the weak spots are stable across a second, cross-family judge: 93% of scores fall within plus or minus 1. A five-run test-retest adds a second axis: even the steadiest model swings from 2 to 10 on a crisis metric across identical re-runs, and deepseek-v4-pro returns a different conversation on every run even at temperature 0. Run-to-run reliability is therefore a per-model safety property, not noise to average away. EMPATH is system-agnostic; the pipeline, seeds, personas, and rubrics are released for reuse.

PromptGNN-sim: Deep Fusion and Alignment of GNN and LLMs for Text-Attributed Graph Learning

Text-Attributed Graphs (TAGs) combine textual semantics with graph structure and are central to many graph learning tasks. However, existing fusion methods often treat text and structure as separate inputs in a shallow, one-way pipeline, which limits deep interaction between modalities and weakens performance under sparse connectivity or cross-graph generalisation. To address this issue, we propose PromptGNN-sim, a bi-directional structure-semantic fusion framework for collaborative GNN-LLM learning. PromptGNN-sim uses a Graph Attention Network (GAT) for semantically aware neighborhood selection by combining structural attention with textual similarity. The selected structural context is then used to generate structure-aware prompts for an LLM, including the target node summary, label categories, and representative keywords from similar neighbors. During training, bi-directional cross-modal contrastive learning and cross-attention are introduced to jointly optimize the GNN and LLM components. Experiments on six public datasets, including Cora, Pubmed, and WikiCS, evaluate accuracy, generalisation, and robustness under cross-task transfer, cross-dataset generalisation, and sparse perturbations. Results show that PromptGNN-sim outperforms classical GNNs, LLMs, and recent GNN-LLM fusion methods, demonstrating the effectiveness of interactive structure-semantic collaboration for text-attributed graph learning.

Rehearsed Multi-Agent Live Product Demonstrations with Real-Time Voice Question Answering

Live product demonstrations are a recurring, high-cost activity in software organizations: a human presenter must select features, dispatch the corresponding interactions on a running application, narrate them coherently, and answer questions in real time. Existing automation addresses only fragments -- generalist browser agents target instruction-conditioned task completion, and demo-video tools produce fixed MP4 artifacts that cannot be questioned and silently break under interface drift. We propose Rhetor, a multi-agent system that takes a running web application and its source-code repository as input and produces a rehearsed live demonstration with segment-synchronized narration and real-time voice question answering. The architectural contributions are a cross-modal feature representation that merges UI exploration with source-code analysis into features tagged with discrete focus tiers, a grounded scripter constrained to UI elements observed during exploration and dispatched through multi-strategy semantic locators, a pre-presentation rehearsal loop with explicit convergence and graceful degradation to narration-only segments, and a runtime synchronization invariant that ties each browser action to the audio-end event of its narration segment. Across six pipeline sessions on four deployed applications -- including the public-domain whiteboard application Excalidraw -- the rehearser's internal locator-firing rate (sigma-bar) spans 0.31-1.00 over 147 scripted actions; on the substantial workload (53 actions, full tier differentiation), sigma-bar is approximately 0.92, and on the public-domain reference point the locator-repair step drives convergence to sigma-bar = 1.00 at iteration 2. We additionally define a benchmark protocol of ten metrics across six application categories that would establish, beyond the case study, whether each design choice contributes positively.

ManimAgent: Self-Evolving Multimodal Agents for Visual Education

Multi-round reflection lets agents built on large language models recover from failures within a single task, but each task remains an isolated episode: lessons learned across many reflection rounds on one task are discarded before the next begins. We study this gap on a code-generation task: from a scientific paper section, the agent writes Python in the open-source Manim library to render a mathematical animation. We present ManimAgent, a self-evolving multimodal agent that carries reflection experience across tasks through a dual-channel Episodic Memory Bank grown entirely from its own task stream, with no weight updates and no human seeds. After each animation converges, a vision-language model scores the rendered keyframes; the resulting signals populate a positive channel M+ that stores success rationales as soft Reference Examples, and a negative channel M- that stores validated failure patterns as hard Known Pitfalls. On a fixed-probe evaluation against no-memory, matched-budget retrieval-augmented generation, and shuffled-memory baselines, blind human Pass@1 rises and reflection rounds fall as memory size grows. We will release the code, frozen memory snapshots, and the task stream.

BayesEvolve: Explicit Belief States for Autonomous Scientific Discovery

Autonomous scientific discovery systems increasingly use large language models (LLMs) to propose new hypotheses, but many such systems condition primarily on experimental memory: archives of high-scoring candidates or heuristic summaries of recent trials. We argue that discovery agents should instead maintain explicit, uncertainty-aware beliefs about hypothesis quality. We introduce BayesEvolve, a belief-guided discovery framework that converts experimental evidence into a predictive belief state and uses this belief to guide future experimentation. As a controlled testbed for belief-guided discovery, we evaluate BayesEvolve on shifted BBOB-style black-box optimization tasks, leaving program and laboratory discovery domains to future work. BayesEvolve improves sample efficiency over memory- and archive-guided LLM baselines under a fixed evaluation budget. We further show that the belief state is predictive on held-out candidate pools, that controlled decision-rule ablations favor belief-guided selection with an annealed uncertainty bonus, and that BayesEvolve exhibits productive late-stage concentration rather than unfocused exploration.

QED: An Open-Source Multi-Agent System for Generating Mathematical Proofs on Open Problems

We present QED, an open-source multi-agent system that turns human-provided research questions into complete mathematical proofs without further human guidance. Its pipeline is designed to overcome common failures of single-query proof generation by separating planning, proving, and verification: a decomposition agent structures the proof search, prover agents generate candidate arguments, and verifier agents check correctness. In collaboration with domain experts, we evaluated QED on 18 research-level projects of varying difficulty. QED produced five original works across algebraic geometry, fluid PDEs, probability, and inverse problems. Expert assessments regard these works as solid specialized research contributions, with three comparable in difficulty and scope to work commonly published in established specialist mathematics venues. QED is released at https://github.com/proofQED/QED.

Geo-Expert: Towards Expert-Level Geological Reasoning via Parameter-Efficient Fine-Tuning

While general-purpose Large Language Models (LLMs) applied to Geology often hallucinate when reasoning about subsurface structures and deep-time evolution, current AI in Earth sciences predominantly targets surface remote sensing and GIS. To bridge this gap, we introduce Geo-Expert, a family of parameter-efficient geological LLMs fine-tuned on a custom-curated, high-quality instruction dataset processed using our custom instruction synthesis pipeline. We investigate the impact of model scaling and architecture by fine-tuning three base models: Qwen3-8B, Qwen3-32B, and Gemma-3-27B, with Low-Rank Adaptation (LoRA) method. Our extensive evaluation on a novel domain-specific benchmark, Geo-Eval, reveals that a domain-aligned 8B model can outperform open-weight 70B generalists and proprietary GPT-4o on specialized geological reasoning, while a 32B variant approaches frontier reasoning models. The optimized 8B model further offers a competitive cost-performance ratio for deployment. This work provides a reproducible recipe for democratizing scientific LLMs and establishes a baseline for geological artificial intelligence.

Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models

Large language models (LLMs) increasingly perform multi-step reasoning, where intermediate claims form implicit directed acyclic graphs whose node correctness is structurally conditioned on their ancestors. This makes factuality uncertainty structural, rather than a trivial accumulation of node-wise errors, and necessitates inference-time uncertainty quantification over the reasoning structure. While conformal prediction (CP) offers flexible user-specified factuality control, existing work remains post-hoc and cannot intervene during generation. To fill the gap between CP's flexibility and its post-hoc limitation, we propose an \emph{Inference-Time Conformal Reasoning (ITCR)} framework that integrates CP directly into reasoning graph generation. ITCR learns a structure-level factuality uncertainty function that aggregates claim-level factuality signals over reasoning graphs without complex modeling assumptions. We then design the non-conformity score based on graph-level factuality uncertainty and calibrate the conformal threshold to decide when to stop generation. We theoretically show such generation is nested, yielding valid coverage guarantees for factuality control. Experiments over multiple datasets and coverage objectives demonstrate empirically valid coverage. In downstream reasoning tasks, inference-time calibrated graphs yield more accurate generation than post-hoc pruned graphs.

Business World Model

World model has emerged as a powerful paradigm in artificial intelligence, enabling agents to represent their environments, predict future states, and evaluate possible actions before acting. However, existing world model approaches have largely been developed for domains such as computer vision, robotics, gaming, and autonomous driving, where the world is primarily visual or physical and governed by relatively stable dynamics. These formulations are not directly applicable to business practice, where the relevant environment is semantic, organizational, and market-driven rather than physical. Business outcomes depend on context-sensitive factors such as customer behavior, pricing, competition, regulation, resources, and operational constraints. This paper introduces the concept and architecture of a Business World Model (BWM), which is a world model specialized for business and organizational environments. A BWM encodes business states, dynamics, and feasible actions space to support autonomous business planning and decision-making. We propose a business-semantics-centric formulation in which states, dynamics, and actions are linked to key business entities, their attributes, and their relationships. Within this framework, intelligent agents can simulate alternative action sequences, estimate their effects on future business outcomes, and evaluate trade-offs under uncertainty. The proposed architecture integrates semantic data representations, probabilistic machine learning models, deterministic business rules, and explicit action spaces into a coherent internal simulator. This work establishes a conceptual foundation for autonomous business systems capable of moving from instruction-based execution toward goal-driven planning, optimization, and execution.

Agents-K1: Towards Agent-native Knowledge Orchestration

Current LLM-based research agents have advanced through agent orchestration, yet largely overlook scientific knowledge orchestration. Existing works often reduce papers to abstracts, surface mentions, and flat \texttt{cites} edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning. To this end, we introduce \textbf{Agents-K1}, an end-to-end knowledge orchestration pipeline that converts raw documents into agent-native scientific knowledge graphs. Agents-K1 integrates three components under a unifying theoretical foundation: a multimodal parser whose five-module schema captures entities, multimodal evidence, citations, and typed inter-entity relations across the full paper rather than abstracts alone; a 4B information-extraction backbone trained with GRPO under a rule-based reward; and a graphanything CLI, a tri-source agent interface that unifies web search, multimodal graph retrieval, and cross-document traversal. On top of this, we process 2.46 million scientific papers across six subjects to produce \textbf{Scholar-KG}, of which we release a one-million-paper subset, and the full Scholar-KG is accessible via the SCP link below. The same pipeline can be extended to general-domain corpora and to schema-conformant data synthesis. Extensive experiments demonstrate that Agents-K1 achieves superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning.

Artificial Intelligence Index Report 2026

Welcome to the ninth edition of the AI Index report. As AI continues to advance rapidly, the question becomes whether the systems built around it can keep up. Governance frameworks, evaluation methods, education systems, and the data infrastructure needed to track AI's impact are struggling to match the pace of the technology itself. That gap between what AI can do and how prepared we are to manage it runs through every chapter of this year's report. New in this edition, the report tracks how AI is being tested more ambitiously across reasoning, safety, and real-world task execution, and why those measurements are increasingly difficult to rely on. It also features new estimates of generative AI's economic value alongside emerging evidence of its labor market effects, an analytical framework on AI sovereignty, and a science chapter developed in collaboration with Schmidt Sciences. For the first time, the report features standalone chapters on AI in science and AI in medicine, reflecting AI's growing impact across these two domains.

RoboPIN: Grounded Embodied Reasoning via Pinned Chain-of-Thought

Embodied reasoning requires models to perceive task-relevant objects and spaces in physical environments and maintain consistent visual grounding throughout multi-step reasoning. However, current vision-language models rely on text-only or coordinate-augmented chain-of-thought, where entity references remain implicit and ambiguous. This may cause the reasoning process to decouple from visual evidence, entity references to drift across steps, and a causal disconnection between the reasoning trajectory and the final answer, with these problems further amplified in multi-view scenarios due to cross-view appearance changes. To address these issues, we propose Pinned Chain-of-Thought (PinCoT), a structured reasoning paradigm that pins every reasoning step to visual evidence. PinCoT introduces the concept of reasoning anchor, which binds each task-relevant entity to a structured visual anchor with entity name, unique identity, view index, and spatial grounding, enabling consistent entity tracking across reasoning steps and views. We build a fully automated data generation pipeline to construct PIN-170K, a high-quality PinCoT-formatted reasoning dataset. We then train RoboPIN through three-stage post-training that progressively injects embodied knowledge, structured reasoning ability, and process-supervised alignment, with rewards that directly constrain both anchor localization and identity consistency during reasoning. On 14 benchmarks covering embodied spatial reasoning, multi-view reasoning, and pointing, RoboPIN with only 4B parameters consistently outperforms 7B level open-source embodied models, achieving a 12% average improvement over the strongest 7B baseline, Mimo-Embodied. Further analysis shows that PinCoT improves grounding accuracy and cross-step identity consistency, validating the effectiveness of process supervision.

Looking Is Not Picking: An Attention-Segment Account of Tool-Selection Failures in LLM Agents

LLM agents mis-call tools, and the natural guess is that the model failed to see the right tool in a crowded harness. We show the opposite through a lens concurrent work sets aside -- the model's attention to labeled tool-definition segments. On real BFCL failures, by per-candidate attention argmax the model attends most to the correct tool 80% of the time (vs. 21% chance), and the gold is the under-attended segment on only 10%: it looks at the right tool and still picks wrong. This directly refutes the intuitive "crowded-harness / lost-in-the-middle" explanation: the failure is at the decision readout, not the harness, and we pin it there three ways. (1) Input vs. readout: repairing the prompt (reordering or duplicating the gold tool) recovers <=23% of failures, while readout-side interventions recover 59-91%. (2) Representation-invariance: two gold-pointed interventions in different representations -- an additive attention-logit bias and a residual-stream steering vector -- recover largely the same failures (per-task Jaccard 0.865 pooled, 0.79-0.91 per model), so the bottleneck is localized to the readout independent of which representation is poked. (3) A training-free, gold-free selector: per-segment attention closes most of the gold-free-vs-oracle gap on BFCL (+11.9 pts pooled function-name selection vs. +17.9-pt oracle headroom) and adds +14.9 pts on Seal-Tools; every model positive (exact McNemar p<=8e-4 each). Scopes differ: the causal attention-bias dose-response is bidirectional and monotonic on 10 mask-honoring models (3-32B), the full 0.5-32B span carrying only the correlational diagnostic; the deployable selector is evaluated on 5 single-turn models and does not yet transfer to a multi-turn loop.

ITNet: A Learnable Integral Transform That Subsumes Convolution, Attention, and Recurrence

Convolutional networks, recurrent networks, and transformers each encode different inductive biases -- locality, sequential memory, and content-dependent pairwise interaction -- and have remained mathematically distinct since their inception. We show that this fragmentation reflects not a fundamental diversity in how signals should be processed, but rather incomplete views of a single underlying mathematical object: a learnable integral transform. We introduce the Integral Transform Network (ITNet), a unified architecture built around a learnable kernel that depends jointly on positions and features. This kernel is implemented as a small neural network, specifically an MLP, that models pairwise interactions, enabling the model to adapt its behavior from data. We show that convolution, self-attention (including multi-head), and autoregressive recurrence (including LSTM, GRU, S4, and Mamba) arise as special cases under appropriate parameterizations, and that ITNet is a universal approximator of continuous operators. To make this practical, we develop tiled kernel fusion, importance-weighted Monte Carlo integration, and learned low-rank factorization, enabling efficient and scalable computation. A single ITNet architecture with a shared operator and lightweight modality-specific encoders matches or exceeds specialized baselines on ImageNet-1K , GLUE, ModelNet40, VQA\,v2 and NLVR2. The results demonstrate that a single learned interaction mechanism can recover the behavior of all three architectural families from data.

The Verification Horizon: No Silver Bullet for Coding Agent Rewards

A classical intuition holds that verifying a solution is easier than producing one. For today's coding agents, this intuition is being inverted: as foundation models develop stronger reasoning capabilities and engineering harnesses grow more sophisticated, generating complex candidate solutions is no longer difficult -- reliably verifying them has become the harder problem. Every verifier we can build is only a proxy for human intent, never the intent itself. This makes verification subject to a twofold difficulty: first, intent is underspecified by nature, making it inherently hard to faithfully check whether it has been fulfilled; second, during model training, optimization widens the gap between proxy and intent -- manifesting as reward hacking or signal saturation. To address this, we characterize the quality of verification signals along three dimensions -- scalability, faithfulness, and robustness -- and argue that achieving all three simultaneously is the central challenge. We further study four reward constructions: a test verifier for general coding tasks, a rubric verifier for frontend tasks, the user as verifier for real-world agent tasks, and an automated agent verifier for long-horizon tasks. Across different task types and policy capability levels, we conduct in-depth analysis and experiments on the core challenges of reward design and how to more effectively leverage reward signals. Experiments show that targeted verification design can effectively suppress reward hacking, improve task completion quality, and achieve significant gains across multiple internal and public benchmarks. These experiences collectively point to a core observation: no fixed reward function can remain effective as policy capability continues to grow; and verification must co-evolve with the generator.

Unbiased Canonical Set-Valued Oracles Via Lattice Theory

A non-agentic "oracle" that reports probabilities of future events is performative: once its answer is learned and acted upon, it can change the very probability it was asked to report. Performativity is not in itself the difficulty -- one consults an oracle precisely in order to be informed, and hence influenced, by it. The difficulty is agency. The requirement that a report be self-consistent, still holding once announced, may be met by many different values -- the classical non-uniqueness of self-fulfilling prophecies -- and any rule the system uses to choose among them is a lever for goal-directed steering. We remove the choice rather than the performativity. Reporting a credal set instead of a single probability distribution, we lift the reaction to an isotone operator on the complete lattice of closed credal sets, whose fixed points are self-consistent, and report its Knaster--Tarski least fixed point as a canonical, rule-determined answer; a variant reports instead the least fixed point that contains every self-consistent point estimate. We prove existence, self-consistency, and nonemptiness; show that the construction reduces to the classical point answer when the question is non-performative; and show that for a binary event the answer is, under a natural hull-factoring assumption, an interval.

JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications

JD$.$com, one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a "Semantic Search then Discrimination"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.