NeurIPS 2026 论文列表

Clinical research over electronic health records depends on derived variables whose definitions sit outside the schema, in item codes, time windows, missingness rules, and output-grain choices that change cohorts and downstream conclusions. We introduce M4Bench, 28 clinical derivation tasks across 15 task families on MIMIC-IV and eICU, with controls that decompose the value of procedural context for AI coding agents into specification completeness, delivery format, and schema familiarity. A Docker-isolated Codex campaign isolates specification completeness as the primary driver: an operational spec over the schema-only baseline yields +0.162 [0.072, 0.260] on eight sentinel tasks, while clinician-reviewed skills add only +0.024 [-0.034, 0.099] on top, and the combined targeted-skill effect over no-skill is +0.136 [0.085, 0.191] across all 28 tasks (26/28 improving). Only 87/760 primary runs pass all strict diagnostics, with failures dominated by output-grain and temporal-boundary errors, and key F1 moves only +0.003 task-balanced, so the gain is partial reference agreement on matched keys, not autonomous task completion or row-grain repair. Schema restructuring preserves 0.704 of the with-skill gain on gpt-5.4-mini over the 26 MIMIC-IV tasks, bounding schema-name memorization at ≤ 29.6% of the effect; decoy-skill and raw-SQL controls further rule out generic-context and reference-code-structure-only explanations. Skills are released as a versioned, clinician-reviewable markdown delivery format, one packaging among several, with task-level reference-validity audit and per-run egress logs. M4Bench releases tasks, skills, evaluation code, and per-run audit logs as a fully open-source benchmark object designed for extension and reuse.

Sparsity has long been a central theme in LLM efficiency but its role in context processing remains unresolved. As LLM workloads shift toward longer contexts and agentic interactions, the compute and memory bottlenecks of attention become increasingly critical, raising the question of whether these constraints are fundamental. Our position: this constraint is artificial, unnecessary and the future of LLM inference lies in extreme but principled sparsity along the context dimension. Our position comes out of various bits of empirical and theoretical evidence. Firstly, we find insistance on dense attention unreasonable since in long context a query essentially projects the O(N) attention information into hidden space of diemsion $d << N$. This, as we show is extremely lossy. Instead, we propose a completely context-sparse context processing combining top-k sparsity with linear attention / SSMs as the way forward. To support our proposal we show that this combination can approximately cover the functional space of dense softmax attention. Independently but importantly, in a first elaborate study of its kind, we empirically show a strong trend towards current LLM models, which are not trained for context sparsity, being extremely robust to inference time decode-sparsity across tasks of varying complexities such as retrieval, multi-hop QA, and mathematical reasoning and agentic coding. For instance, Qwen3.5-27B can tolerate upto 100x sparsity on benchmarks of RULER-HARD, LOFT and AIME2025 without loss of quality, and upto $50\times$ on SWE with a small drop in quality. These results emphasize the possibility that we can transition to complete sparsity without any loss of capability. Additionally, we also discuss what this shift in paradigm of context processing means for hardware. Importantly, we show that even current hardware is equipped enough to realise gains from this sparsity. For instance, our sparse decode kernels can accelerate large context processing by a factor 10x over FlashInfer at 50x sparsity levels on current hardware such as H100. Overall, these results position extreme context sparsity not as a heuristic, but as a principled foundation for LLM inference, training, and architecture design, both feasible and beneficial, and a compelling direction for future systems.

Long-form generation in language models often collapses into repetitive loops well before the nominal context limit. This is commonly attributed to model capacity or training data, and addressed through costly retraining or post-training pipelines. We show that the sampling algorithm is a primary determinant of long-context output quality, improving outputs without further training costs. Across ten open-weight models from 1B to 120B parameters, output lengths from 8K to 64K tokens, and more than 13,500 generations, full-distribution-aware samplers (P-less, Top-H, top-$n\sigma$) dominate automated metrics and human evaluation, while standard truncation (top-$p$, top-$k$) collapses into late-context repetition; min-$p$ falls in a middle tier. A 1,500-annotation Amazon Mechanical Turk study finds P-less winning more than nine in ten head-to-head pairings against a top-$p$ 0.9 baseline on a 120B model, with 10-gram repetition approaching zero at 64K tokens. A per-step selected-token rank diagnostic isolates the mechanism: standard truncation drives the rank to the argmax floor by the final quartile, while full-distribution-aware methods stay in a healthy band across the full window. This gap that widens with model scale, establishing sampler choice as an under-appreciated bottleneck for long-context generation in open-weight models. We frame long-context generation as sampling from outside the training distribution, and encourage further study for usecases like long-context agentic coding.

Multimodal large language models (MLLMs) have made rapid progress in general visual understanding, yet they still struggle with fine-grained instance-level consistency verification, where the goal is to determine whether two images depict the exact same physical object under varying viewpoints, illumination, and backgrounds. This task requires both a global understanding of the target instance and a precise examination of subtle local evidence, since semantically similar objects may differ only in small discriminative regions. To address this challenge, we propose \textbf{LocalAgent}, a collaborative agentic framework that integrates the complementary strengths of large and small models. Specifically, the MLLM first reasons from a global instance-level perspective to identify suspicious local regions that may determine the final verification result. These accurately localized regions are then delegated to a lightweight Local Evidence Verifier (LEV) toolkit for fine-grained appearance and geometric matching. The LEV toolkit consists of \textbf{Local Appearance Consistency (LAC)}, which provides illumination- and color-robust similarity estimation, and \textbf{Local Geometric Correspondence (LGC)}, which establishes reliable spatial correspondence under viewpoint changes. Rather than directly replacing the MLLM's reasoning, LEV returns structured similarity scores and comparative evidence in a form that the MLLM can effectively interpret for final decision-making. This design enables a coordinated large-small model collaboration: the MLLM contributes global semantic reasoning and region-level suspicion localization, while specialized local verifiers provide precise, invariant, and interpretable evidence for ambiguous cases. We further introduce an answer-dominant reinforcement fine-tuning (RFT) strategy that enables the policy to learn when verification is necessary. Experiments show that LocalAgent achieves a \textbf{+4.8\%} average improvement over direct-inference baselines, effectively narrowing the gap with closed-source models, while preserving the general multimodal capabilities of the underlying MLLM.

Large language models (LLMs) and multi-agent systems (MAS) have shown promise in financial decision-making, yet existing evaluations focus on equity trading and primarily assess directional prediction, overlooking the structural complexity of derivative markets. Option trading introduces fundamentally different challenges, including nonlinear payoffs and multi-leg strategy construction, requiring structured decisions rather than simple directional bets. We introduce LiveOption, an evaluation framework for LLM-based and reinforcement learning agents in option trading. LiveOption formulates the problem as structured sequential decision-making under realistic execution and capital constraints, and provides a reproducible environment with standardized interaction protocols. The framework includes several task suites covering portfolio overlays, open-ended strategy construction, event-driven earnings trading, and 0DTE intraday trading. We further propose a hierarchical metric suite that evaluates action validity, decision quality, risk characteristics, and outcome-level performance. Experiments show that current agents often fail to construct valid strategies and handle nonlinear payoff dynamics, or achieve competitive returns in most scenarios. LiveOption offers a principled testbed for evaluating structured decision-making beyond outcome-based metrics.

As AI agents are given more autonomy in software engineering workflows, the risks grow if they pursue goals different from the ones their users intended. The field of AI control develops control protocols that prevent such harm without restricting the agent's ability to do useful work. We introduce \textbf{LinuxArena}, a control setting for testing such protocols, where agents operate directly on live, multi-service production environments. LinuxArena contains 10 public environments (with an additional 10 held out privately), providing 906 main tasks representing legitimate software engineering work and 92 side tasks representing safety failures such as data exfiltration and access-control bypassing, making it the largest and most diverse control setting for agentic software engineering to date. We demonstrate LinuxArena's utility by running sabotage and monitor evaluations: against a GPT-5 Nano trusted monitor at a 1% step-wise false positive rate, Claude Opus 4.6 achieves an undetected sabotage rate of 42%. We additionally release \textbf{LaStraj}, a dataset of human-crafted attack trajectories that achieves an undetected sabotage rate of 97% against the same monitor. Together, these results suggest meaningful headroom for both attackers and defenders, making LinuxArena a strong testbed for developing and evaluating future control protocols.

Large language model agents now combine tools, long context, memory, and multi-step workflows, yet they still fail when execution loses track of the evidence required for a reliable answer. Common failure modes include missing support, unresolved contradictions, and evidence lost under context pressure. We present Lemon, an evidence-risk-aware framework that makes such risks an explicit control variable. At the intra-instance level, a single state-conditioned controller jointly schedules reasoning, tool invocation, worker expansion, verification, context compression, and memory operations. A recoverable evidence substrate pairs anchored context compression with reusable semantic memory, preserving raw tool outputs while storing transferable process fragments. We further extend Lemon to an inter-instance setting in which heterogeneous personalized agents expose expertise, exchange evidence-bearing proposals, critique unsupported claims, and internalize useful collaboration artifacts. Empirically, Lemon reaches 91.36\% accuracy on GAIA and 80\% on xbench-DeepSearch. In an open-source reproducibility comparison on GAIA, Lemon uses 49.9\% to 88.4\% fewer tokens per task than three top-ranked agent baselines.

Training tool-using agents for planning is challenging when high-quality demonstrations exist only as final outcomes without the intermediate tool traces that produced them. A common response is to impute traces and apply imitation learning, but this approach is indirect and structurally limited to matching the reference rather than improving upon it. We propose anchored comparative feedback: rather than imitating a reference plan, we train agents to surpass it under a comparative LLM judge while satisfying hard constraints. The anchor provides a stable optimization target, and comparing two identically rendered plans neutralizes surface-level judge biases. On TripTailor, a tool-augmented travel planning benchmark, anchored GRPO achieves 80.1% final success rate compared to 43.7% for imitation learning on imputed traces, with gains on both constraint satisfaction and LLM-judged preference that persist across judge families, rubrics, and a held-out reward model. Gains generalize beyond TripTailor, with a fivefold improvement on the held-out TravelPlanner benchmark. These results suggest that anchored comparative feedback offers an effective approach for learning transferable tool-using planners from outcome-only supervision.

Persuasion is a core dynamic of natural language communication, shaping how large language models (LLMs) update beliefs, resolve disagreements, and reach decisions. As LLMs increasingly debate, advise, and think collaboratively with humans and each other, resistance to harmful persuasion becomes a core requirement for reliable behavior. Yet we show that this requirement is far from met: a single targeted persuasive argument is enough to collapse model accuracy to near zero, even when the argument is factually false. We formalize this threat as adversarial persuasion and introduce an adversarial reinforcement learning framework that trains persuader agents to change a target model's answer in a single interaction. First, we show that optimizing persuasion strategies through trial and error exposes vulnerabilities that static prompting misses: RL-trained persuaders raise persuasion success from approximately 24\% to over 93\% against the training-time persuadee. Second, we find that these learned strategies transfer to unseen models, achieving 83\% attack success on Qwen-14B, 79\% on Llama-3.1-8B, and 25\% on GPT-4o-mini. Third, we demonstrate that a curriculum that bootstraps on more persuadable open-weight models before targeting harder models further increases GPT-4o-mini attack success from 25\% to 38\%. Moreover, our results reveal that optimized persuaders increasingly rely on credibility-based tactics, including fabricated citations and false authoritative evidence. Together, these findings expose a critical weakness in current LLM agents: even when they initially reason correctly, they can be steered toward false conclusions by optimized natural language influence. This positions persuasion robustness as a necessary safety criterion for multi-agent and human-AI decision-making systems.

Large language model (LLM)-based coding agents achieve strong results on controlled benchmarks yet routinely produce pull requests that real maintainers reject. The cause is not functional incorrectness but a lack of organicity: generated code ignores project conventions, duplicates internal APIs, and violates implicit architectural constraints. The latest repository snapshot is insufficient-it reveals the final codebase state, not the change patterns that shaped it. We make three contributions. (1) Paradigm: we propose next-commit prediction as a paradigm for learning agentic coding skill from a repository's own history-each historical commit is a self-supervised target whose oracle diff supplies dense supervision. (2) Evaluation: we establish Organicity Evaluation as a measurable objective for coding agents and contribute the Learning to Commit benchmark-5 curated open-source GitHub repositories under strict per-repository temporal splits-scoring patches along file localisation, internal API reuse, patch bloat, and code-style consistency. (3) Method: we instantiate the paradigm with the Learning to Commit framework, in which the agent performs supervised contrastive reflection: it blindly attempts each historical commit, contrasts its prediction against the oracle, and incrementally distils a reusable skill document that conditions subsequent generations. On the Learning to Commit benchmark and on SWE-bench Pro reframed under our temporal-split protocol, our framework consistently improves organicity on held-out future tasks and further lifts test-pass rates on SWE-bench Pro, narrowing the gap between benchmark success and real-world mergeability.

Markov Games are the standard formal model for multi-agent reinforcement learning, capturing agents that act in a shared state and optimize rewards over time. However, in real-world settings, agents’ decisions are often influenced by unobserved factors, such as cognitive biases, behavioral tendencies, or intuitive signals, that also affect rewards and future states. Ignoring these unobserved confounders can make standard learning methods converge to suboptimal policies. In such settings, optimal play may require policies that condition on counterfactual signals, requiring reasoning at the counterfactual layer of the Pearl Causal Hierarchy. In this paper, we introduce Causal Markov Games (CMGs), a framework for modeling sequential multi-agent decision making in the presence of unobserved confounding. We show that CMGs strictly generalize Markov Games, with arbitrarily large gaps between classical interventional equilibria and causal counterparts. We then develop two learning algorithms under different observability assumptions. The first, CNash-VI-FO, is a model-based learner with finite-sample guarantees when agents’ natural actions or intuitions are revealed post hoc. The second, CNash-VI-NO, is an explore-then-exploit procedure with asymptotic guarantees for the setting in which opponents’ natural actions are never observed. To address scalability, we further provide a drop-in counterfactual augmentation of deep MARL. Empirically, on a confounded windy variant of the Multi-Particle Environment and the Iterated Causal Prisoner’s Dilemma, counterfactual agents strictly dominate their non-causal counterparts.

Multi-agent systems have emerged as a promising paradigm for multimodal medical reasoning, but existing methods rely on two assumptions that hinder practical deployment: (i) always-active dense collaboration among multiple agents, and (ii) dependence on large cloud-based models that raise privacy and efficiency concerns. In this paper, we challenge both assumptions and propose SMART-Med, a fully local multi-agent method that treats collaboration as a learnable decision rather than a default behavior. Our key insight is that not every medical query requires costly multi-agent deliberation: many can be reliably solved by a single well-trained agent, while only uncertain cases benefit from expert collaboration. Based on this insight, SMART-Med introduces an uncertainty-driven routing mechanism that adaptively decides when to answer directly and when to recruit a small subset of expert agents, effectively converting dense multi-agent reasoning into selective, query-adaptive coordination. To make this routing reliable, we curate a new multimodal medical dataset MedVR-3K with high-quality reasoning traces and a difficulty hierarchy, and then fine-tune local agents using GRPO with two complementary rewards: a certainty-aware accuracy reward and a vision-language reward model for correctness and reasoning quality respectively. Experiments on four medical visual question answering benchmarks show that SMART-Med achieves competitive performance compared to existing MAS approaches while substantially reducing inference cost (reducing inference time by more than 3.8$\times$ and token usage by 72\%). Moreover, after fine-tuning on MedVR-3K, SMART-Med yields an average improvement of 6.7\% across Slake, PathVQA, VQA-RAD, and PMC-VQA. Our results suggest that selective collaboration can be an efficient yet effective alternative to dense collaboration and enable practical, privacy-preserving deployment of multi-agent medical AI.

Machine unlearning aims to remove targeted knowledge from a trained model while preserving its general capabilities. Doing so effectively for auto-regressive language models requires identifying which tokens in a forget sample are truly relevant to forgetting, as uniformly applying unlearning across all tokens can degrade utility. Existing approaches, however, either ignore this heterogeneity or rely on auxiliary models, hand-crafted heuristics, or external annotations to approximate token-level relevance. We instead characterize token-level forget relevance through its interaction with the retain objective: tokens are forget-relevant to the extent that minimizing the forget loss on them does not conflict with the retain objective. Based on this insight, we introduce Alternating Token-Weighted Unlearning (ATWU), a framework that jointly learns token relevance and model parameters during the unlearning process. ATWU uses a lightweight linear scorer over model hidden states to predict forget-relevance scores with negligible computational overhead and no external supervision. Experiments on TOFU and RWKU show that ATWU achieves state-of-the-art forget--retain trade-offs, outperforming sample-level methods, probability-based token-weighting heuristics, and auxiliary-model-based approaches. Moreover, the learned scorer aligns with ground-truth forget-relevant spans substantially better than existing methods, suggesting that ATWU learns semantically meaningful token-level forgetting signals. Overall, ATWU shows that token-level forget-relevance can be effectively inferred from model representations during unlearning, enabling efficient and selective forgetting.

Recent work has identified incremental learning in shallow networks trained on single-index and multi-index models. However, existing analyses often rely on simplifying settings, such as small initialization, correlation loss, or layer-wise training. These choices reduce neuron interactions and leave some feature learning dynamics under standard initialization unexplored. We study gradient flow dynamics for polynomial-width two-layer networks learning orthogonal multi-index targets under standard initialization using polynomially many samples. We first prove that incremental learning still occurs: the loss decreases sequentially according to the Hermite expansion of the target, with lower-order components learned before higher-order components recover the individual target directions. In this standard initialization regime, training also shows a competitive reallocation of parameter mass: after the total mass fits the target mean and stabilizes, mass shifts into the target subspace and then concentrates on aligned neurons. Technically, we introduce a symmetry-based finite-width approximation via symmetrized networks, rather than comparing directly with an infinite-width limit. This yields better control of approximation errors and may be of independent interest.

While multi-agent reinforcement learning (MARL) has shown strong empirical performance, existing methods struggle with large number of agents, due to the combinatorial growth of joint interactions. Mean-field (MF) approximations address this by replacing pairwise interactions with interactions against population distributions, yielding tractable large-population policies. However, existing MF formulations focus on fully cooperative or purely competitive settings and fail to capture the mixed cooperative–competitive structure of team-based games. We extend mean-field learning to _zero-sum team games_, where agents cooperate within teams and compete at the team level. We show that such games admit $\epsilon$-optimal decentralized policies that depend only on local states and population distributions. Building on this structure, we propose MF-MAPPO, a scalable algorithm with a shared actor and a minimally informed critic per team. MF-MAPPO is trained directly in finite-population simulators rather than using mean-field oracles, thereby enabling deployment to realistic scenarios with thousands of agents. We further extend MF-MAPPO to partially observable settings via a simple gradient-regularized training scheme. Experiments on large-scale benchmarks in our simulation platform $\texttt{MFEnv}$, including population games with analytical solutions and high-dimensional battlefield scenarios, demonstrate that MF-MAPPO outperforms existing MARL baselines and yields rich, heterogeneous behaviors.

We study online identification of linear-in-parameter dynamical systems whose true parameters drift over time. Measuring non-stationarity through the path variation $V_T = \sum_{t=1}^{T-1}\|\theta_{t+1}^\star - \theta_t^\star\|$, we make two contributions. First, we give a clean blockwise dynamic-regret analysis for continuous projected online gradient descent, yielding $E[R_T] = O(T/\sqrt{\tilde W} + \tilde W V_T)$ and hence the oracle-tuned rate $O(T^{2/3}V_T^{1/3})$ . Second, we propose a residual-based adaptive step size and prove an oracle-proxy comparison theorem: when the residual proxy tracks the local prediction-error signal with lower-order overhead, the adaptive algorithm inherits the oracle scaling. Experiments on synthetic systems and real turbofan-engine degradation data (NASA C-MAPSS) support the matched-regime slope prediction on synthetic data, diagnose the failure regime when proxy overhead dominates, and show a clear unsupervised health-indicator trend in the residual proxy on FD001. A separate fast-rate guarantee under persistent excitation and a tube MPC integration are deferred to the appendix.

Command line interface (CLI) agents are emerging as a practical paradigm for agent-computer interaction over evolving filesystems, executable command line programs, and online execution feedback. Recent work has used reinforcement learning (RL) to learn these interaction abilities from verifiable task feedback, yet few methods exploit the native structured attributes of CLI actions as learning signals. Beyond this underused action structure, CLI learning also couples two bottlenecks for coding agents. First, the agent must identify task-relevant evidence in a large codebase from partial observations. Second, sparse terminal rewards must be assigned to the actions that shape a long multi-turn trajectory. We study these bottlenecks through shell-driven information extraction and file editing tasks. For selective observation, we introduce $\sigma$-Reveal, an inference-time mechanism that selects token-budgeted context for the same CLI. For credit assignment, we propose Action Advantage Assignment ($\mathrm{A}^3$), a native agentic RL method that preserves the algorithmic complexity of standard agentic RL. $\mathrm{A}^3$ constructs turn-level advantages from episode-level relative feedback, abstract syntax tree (AST) based action sub-chain residuals, and tree-level trajectory margins. To further evaluate this problem setting, we construct ShellOps, a verifiable dataset suite covering CLI tasks in repository environments.

Recent progress in integrating vision-language-action (VLA) models with world modeling has significantly advanced end-to-end autonomous driving by allowing systems to reason about imagined futures instead of merely reacting to current observations. However, most world VLA models generate future frames at the pixel level with plenty of dense visual tokens. Although such generated futures can be visually plausible, they preserve many driving-irrelevant appearance details, i.e., only a small portion of the tokens directly encode planning-relevant factors, such as agent geometry, motion dynamics, and interactions. This introduces a significant performance bottleneck. An important question thus arises: what world state should a world VLA model reason over to support planning beyond pixel-level prediction? We argue that planning-oriented world modeling should focus on an agentic latent state that captures the interactions between objects in the driving scene over time. As such, we introduce AgentWorld, an agentic world VLA model that uses discrete agent tokens jointly for reasoning and trajectory planning. Each token is dynamically grounded to a surrounding agent associated with its object's state, such as semantic identity, 3D geometry, motion, and interaction context. AgentWorld learns these tokens through three key processes: explicit geometric grounding, reasoning-oriented supervised fine-tuning, and agent-oriented reinforcement learning. Extensive open-loop and closed-loop experiments under diverse driving scenarios show that AgentWorld enhances planning capabilities, fully exploits visual evidence and agent dynamics, and yields interpretable agentic world states. The code will be made publicly available.

Reasoning, planning, and agentic systems demand that neural models extrapolate from the easy training instances to harder, more compositional problem instances they encounter at deployment. Neural networks trained by gradient descent are reliable interpolators \emph{within} their training domain but struggle to extrapolate beyond it: they learn local generalizers even when global solutions are representable in their architecture class. We propose a \emph{learn locally, recurse globally} paradigm-: train a neural model on in-distribution sub-problems and apply it iteratively under a verifiable outer loop that decomposes hard instances into a sequence of in-distribution sub-instances and reassembles the result. We instantiate this paradigm in Boolean circuit synthesis from truth-tables, which we adopt as a faithful, controllable abstraction of logical reasoning: any propositional inference reduces to evaluating a Boolean circuit, and circuit \emph{depth} is a clean proxy for the difficulty of the underlying reasoning task. Empirically, a $2.2$M-parameter encoder-decoder Transformer trained on depth-$\leq 2$ circuits, wrapped in our outer loops, synthesizes circuits at $5\times$ training depth and at input arities for which the base model's encoder is too small for direct inference, with synthesis quality matching a standard EDA tool while producing markedly shallower circuits. We connect the algorithms to classical Boolean decomposition theory (Shannon, Ashenhurst-Curtis, Reed-Muller) and to recent formal results on the AC$^0$ barrier for constant-depth Transformers.