NeurIPS 2026 论文列表

Language models are increasingly used in decision pipelines where an initial answer is revised after another agent disagrees. In such settings, revision can reflect useful reconsideration, ordinary prompt instability, or sensitivity to the perceived prestige of the disagreeing source. We introduce a controlled two-pass audit that separates these effects by holding the question and evidence fixed while varying only the revision context: neutral reflection, anonymous disagreement, and expert-labeled disagreement. We evaluate five instruction-tuned models on 1,989 evidence-grounded binary decisions from medicine, scientific claim verification, and contract reasoning. Disagreement consistently induces reversals beyond drift, while the incremental effect of an expert label varies across models. More importantly, these reversals are often harmful: under expert-labeled disagreement, 70.1% of reversals aggregated across models move from correct to wrong answers. In a branch-paired API subset, low-prestige disagreement is weaker than expert disagreement, and a simple evidence-gated revision prompt reduces harmful reversals while improving final accuracy. For open-weight models, teacher-forced Yes/No scores reveal that many output reversals are not accompanied by corresponding preference-sign changes. These results make prestige-sensitive revision a concrete, reproducible, and actionable target for LLM evaluation.

LLM powered multi-agent systems (MAS) have emerged as a promising paradigm for complex tasks. However, their advantages over single-agent systems (SAS) remain unclear, with performance varying inconsistently across settings. Here, we provide an information bottleneck perspective on elucidating the differences between MAS and SAS. Specifically, our key observation is that a SAS accumulates its full reasoning trace in one shared context, while a MAS uses isolated local contexts connected by bounded relay messages. We show that, under infinite relay bandwidth, any SAS can be simulated by a MAS that transmits the full upstream context. Thus, the nontrivial advantage of MAS arises under bounded relays, where compression introduces a fundamental trade-off: reducing redundant context can improve efficiency, but may also incur loss of task-relevant information. We formalize this trade-off as an information bottleneck controlled by an effective parameter $\beta$, which captures how the balance shifts with model capability, and shows that MAS gains arise when context reduction outweighs relay information loss. We conduct 18 controlled experiments across five benchmarks and three model scales to validate our theoretical studies. We observe that MAS consistently helps when relays are near-sufficient, especially for weaker models. In contrast, MAS gains shrink or reverse when relays incur information loss, especially for stronger models that can already extract useful information from redundant context and thus gain little from compression. Our study shows that multi-agent design is fundamentally an information-bottleneck optimization problem. This perspective explains when bounded inter-agent communication helps or hurts.

Multi-agent debate can improve reasoning, yet often fails to beat simple majority voting. Prior martingale-null theory explains such failures as debate without truth-directed signal. We develop a more complete account of when practical LLM debate succeeds: it needs both correct minority proposals and a readout that can recover them. We formalize this view through latent headroom, the gap between majority vote and proposal-oracle performance, which measures unused correct expertise in an expert society. For readout, we developLatent Verification Debate (LVD), a theory that models candidate proposals as receiving latent verified evidence before influencing final generation. For supply, we study neural-thicket expert societies, where nearby model perturbations provide heterogeneous specialists, and introduce a coverage-based construction that increases complementary proposal supply. Across matched-budget reasoning and multi-discipline benchmarks, our construction increases recoverable headroom, and controlled readout experiments show that gains arise from recovering surfaced correct minorities rather than merely adding interaction. Together, these results identify diverse proposal supply and verification-aware utilization as the two mechanisms that determine when debate helps.

It is widely believed that stochastic gradient descent (SGD) performs significantly worse than adaptive optimizers such as Adam in pre-training Large Language Models (LLMs). Yet the underlying reason for this gap remains unclear. In this work, we attribute a large part of the discrepancy to SGD's inability to sustain learning rates comparable to Adam's much larger effective learning rates. Through empirical and theoretical analysis of LLM pre-training dynamics, we identify that training is characterized by small gradient norms and large weight-to-gradient ratios, an effect that becomes more pronounced with larger batch sizes typical in pre-training, necessitating such large effective learning rates. However, we find that output-layer gradient magnitudes become highly uneven across token classes, and that large gradient spikes frequently occur during training. Together, these effects severely restrict the admissible learning rate of SGD. Guided by this understanding, we show that simple clipping mechanisms that stabilize SGD at large learning rates enable it to recover most of Adam's performance. In our large-scale experiments, the validation loss gap between large-learning-rate SGD and Adam shrinks from more than 50% to only about 3.5% when pre-training a 1B-parameter LLaMA model with a 1M-token batch size.

LLM unlearning is often implemented by assigning surrogate targets, such as refusal, uniformity, or likelihood suppression, to forget-related prompts. We argue that this surrogate assignment view is misaligned with the counterfactual goal of unlearning: recovering the retain-only predictive posterior. This mismatch is especially problematic in mixed-evidence regimes, where a response may be associated with forgotten evidence while still being partially supported by retained evidence. We propose EASE: Evidence Attribution and Subtraction Estimator, a logit-space posterior-correction method that estimates the forget-induced component of the deployed model's predictive support and subtracts it at inference time. EASE uses lightweight deletion and compensation assistants to remove forget-neighbour evidence while restoring nearby retain-supported evidence. We formalize the suppression-scope dilemma and provide guarantees showing when posterior correction recovers the retain-only posterior. Experiments on TOFU and MUSE show that EASE improves forgetting--retention trade-offs over strong baselines, achieving the best aggregate score in 5/6 main TOFU settings, reducing mixed-query semantic leakage by 22.2%, and matching or exceeding full-retain baselines with 9 times fewer retain examples. Our code is available at: https://anonymous.4open.science/r/EASE-9675.

Training multi-turn LLM agents with reinforcement learning requires distributing a single episode reward across many interaction steps. Recent group-based methods tackle this credit assignment challenge by finding steps where the agent visits the "same" state across different rollouts and comparing their outcomes. However, determining when two states are "the same" remains an open problem: existing methods compare raw text observations via exact string matching or bag-of-words similarity, both of which ignore whether state differences actually matter for future returns. We show that this is a state abstraction problem with a precise bias-variance tradeoff: matching on value-irrelevant features creates unnecessarily small groups (high variance), while ignoring value-relevant features merges states that should be distinguished (high bias). Guided by this analysis, we propose Structured State Abstraction (SSA), a step-level credit assignment method that decomposes observations into typed fields using environment-provided metadata and learns per-field importance weights from return variance. Rather than hard grouping, SSA uses a kernel-smoothed advantage estimator where all states contribute to each step's baseline proportionally to their structured similarity, yielding lower MSE than threshold-based grouping. Experiments on ALFWorld and WebShop show consistent improvements over GiGPO and HGPO across both benchmarks and model scales; notably, kernel smoothing eliminates the singleton problem that causes existing methods to lose step-level credit for 30-36% of steps. Our implementation is provided in the supplementary materials and will be open-sourced upon acceptance.

Reusing a held-out benchmark adaptively should, in principle, invite overfitting. Yet benchmark-driven machine learning (ML) has produced surprisingly little overfitting in practice. An attractive hypothesis is that successful ML strategies are highly compressible. We study this in the setting of LLM-driven research agents, where the hypothesis becomes directly testable via two complementary information bottlenecks. In \emph{output compression}, an exploration agent adaptively searches for high-performance models using a validation set, and we test whether a fresh ``reproducer agent'' can reproduce its performance given only an extremely short prompt and the training data. In \emph{input compression}, the explorer receives only one-bit feedback indicating whether each submitted model improves on the running best. Across 8 datasets spanning tabular classification, vision, language modeling, diffusion modeling, and reward modeling, we find that these bottlenecks have little effect on performance: short prompts and compressible feedback are sufficient to reproduce and find high-performance models. The hypothesis is falsifiable: when we deliberately induce validation-set overfitting, the results fail to reproduce with short prompts. Taken together, our results support a description-length explanation for the lack of overfitting in benchmark-driven ML: successful strategies occupy a low-complexity region of strategy space.

LLM benchmark scores are often read as claims about factual accuracy, evidence use, instruction following, or logical reasoning. We study this score-to-claim conversion as inference over a completed evaluation record. Given benchmark items, prompts, outputs, scores, and native fields, which capability sentence is supported, which stronger sentence remains outside the evidence, and would that boundary change model choice? We formalize a report-use procedure, \emph{score-to-claim inference}: hold source units fixed, compare each claim-induced condition to benchmark-native controls or same-source invariance checks, and apply a declared finite-family support rule. In a fixed set of five open-weight model families, completed records support narrower sentences than endpoint readings alone suggest: supplied-context answer recovery remains separate from support attribution; local supplied-evidence verdict signals leave open-web provenance unmeasured; and SATBench exposes label-surface instability for SAT/UNSAT prediction. The output is a support boundary, a blocked strengthening, and the next measurement or model-selection sensitivity attached to the reported result.

Watermarking techniques for large language models (LLMs), which encode hidden information in the output so its source can be verified, have gained significant attention in recent days, thanks to their potential capability to detect accidental or deliberate misuse. Similar challenges involving model misuse also exist in the context of game-playing, such as when detecting the unauthorized use of AI tools in gaming platforms (e.g., cheating in online chess). In this paper, we initiate the study of how game-playing strategies can be watermarked. We show how the KGW watermark for LLMs can be adapted to watermark game-playing agents in perfect-information extensive-form games. The watermark can then be detected using a statistical test. We show that the degradation in the quality of the watermarked strategy profile, quantified by the expected utility, can be bounded, but there is a tradeoff between detectability and quality. In our experiments, we bootstrap the watermarking framework to various chess engines and demonstrate that a) the impact of the watermark on the quality of the strategy is negligible and b) the watermark can be detected with just a handful of games.

Large language model (LLM) agents increasingly operate through multi-step tool-use trajectories, where harmful intent may be distributed across actions that appear benign in isolation. This challenges single-step moderation, while repeated LLM-as-judge evaluation over growing prefixes can be costly and may intervene too late. We introduce *WURI* (*Watching Unfolding Risk in Agent Interactions*), a lightweight monitor for early prefix-level detection of harmful agent trajectories. WURI encodes each observed textual step with a frozen text encoder, learns an atom-adapted trajectory representation through metric learning, and scores each prefix with a fixed prototype-margin rule. It requires no access to agent model weights or hidden states, does not modify the agent's internal model, and uses no external LLM-as-judge calls at inference. Across five generalization settings, WURI achieves the best average prefix-area under the detection curve (AUDC), ranks first on three settings, and reaches the strict early-detection operating point within four steps in multiple evaluation settings. Runtime analysis shows a wall-clock speedup of over two orders of magnitude compared with repeated full-prefix guardrail evaluation. These results demonstrate that representation-based prefix scoring is an effective and efficient direction for monitoring unfolding risk in LLM agent interactions.

Sparse attention is becoming increasingly important for serving large language models (LLMs) as generation lengths continue to grow. However, deploying and evaluating new sparse attention algorithms at scale remains highly engineering-intensive, slowing both human researchers and AI agents in exploring the sparse attention design. To address this challenge, we present Vortex, a system that combines a Python-embedded frontend language atop a page-centric tensor abstraction for expressing a broad range of sparse attention algorithms, with an efficient backend tightly integrated into modern LLM serving stacks. Vortex enables rapid prototyping, deployment, and evaluation of sparse attention algorithms, effectively translating their theoretical efficiency gains into real-world throughput improvements. As a result, Vortex substantially accelerates the design and iteration of sparse attention algorithms. Using Vortex, we further demonstrate AI-agent-driven exploration of the sparse attention, automatically generating and refining diverse algorithms that achieve strong accuracy-throughput trade-offs, with the best generated algorithms delivering up to $3.46\times$ higher throughput than full attention while preserving accuracy.

Multimodal agents struggle with physical-world reasoning, particularly spatial reasoning from partial views, as the physical world is inherently complex. We think multimodal agentic reasoning about the physical world should ground its actions in deterministic feedback, rather than rely solely on internal imagination. Concretely, we propose to equip vision–language models (VLMs) with a physics engine (e.g., MuJoCo, UE5), which simulates the consequences of the agent's actions. However, granting the agent raw access to a physics engine is not sufficient, as the agent must first reconstruct the scene inside the physics engine. We therefore introduce the Visual Harness System, an orchestration system that decouples perception from reasoning with a physics engine as the backend. The perception module composes off-the-shelf perception skills, such as camera-pose estimation, open-vocabulary detection, and metric depth, into a structured scene map. The reasoning module then issues multi-turn tool calls that the engine executes over the map, and composes the final answer from the grounded feedback it receives. Experiments on spatial reasoning benchmarks show that our approach consistently outperforms strong baselines across both open- and closed-weight VLMs, with up to $16.57\%$ improvement over prior methods on the MindCube benchmark.

Visual content generation has moved from single-image to multi-image workflows, yet existing agents are largely plan-driven and lack a systematic reflection mechanism for correcting mid-trajectory visual errors. To address this gap, we propose VisionCreator-R1, a native visual generation agent with explicit reflection, paired with a Reflection-Plan Co-Optimization (RPCO) training methodology. Through extensive experiments and trajectory-level analysis, we uncover a reflection-plan optimization asymmetry in reinforcement learning (RL): planning can be reliably optimized via plan rewards, while reflection learning is held back by noisy credit assignment. Motivated by this finding, we further abstract a general Decouple-then-Fuse paradigm for co-optimizing capabilities with asymmetric reward variance, of which RPCO is the visual-generation instantiation. Our RPCO first trains on the self-constructed VCR-SFT dataset, which covers reflection-strong single-image trajectories and planning-strong multi-image trajectories, and then co-optimizes on the VCR-RL dataset via RL. This yields our unified VisionCreator-R1 agent, which beats Gemini2.5Pro on existing benchmarks and on our VCR-Bench covering single-image and multi-image tasks.

Multimodal Large Language Models (MLLMs) still struggle with fine-grained visual understanding, where answers often depend on small but decisive evidence in the full image. We observe a regional-to-global perception gap: the same MLLM answers fine-grained questions more accurately when conditioned on evidence-centered crops than on the corresponding full images, suggesting that many failures stem from difficulty to focus on relevant evidence rather than insufficient local recognition ability. Motivated by this observation, we propose Vision-OPD (Vision On-Policy Distillation), a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy. Vision-OPD instantiates two conditional policies from the same MLLM: a crop-conditioned teacher and a full-image-conditioned student. The student generates on-policy rollouts, and Vision-OPD minimizes token-level divergence between the teacher and student next-token distributions along these rollouts. This enables the model to internalize the benefit of visual zooming without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use. Experiments on multiple fine-grained visual understanding benchmarks show that Vision-OPD models achieve competitive or superior performance against much larger open-source, closed-source, and "Thinking-with-Images" agentic models.

Long-form video understanding (LVU) requires models to reason over hour-long narratives while preserving fine-grained spatiotemporal details across extended temporal horizons. Although recent multimodal large language models (MLLMs) have achieved strong performance on short-video tasks, their reasoning capabilities degrade substantially as video duration and information density increase. We present VideoSleuth, a narrative-centric agentic framework for LVU built on video- and audio-native MLLMs. VideoSleuth organizes its memory around the four narrative elements-time, place, person, and event-and equips the agent with tools explicitly designed to construct and query these elements within a ReAct-style reasoning loop. Unlike prior LVU agents that rely on dense offline preprocessing or image-only VLMs, VideoSleuth performs perception on demand using a video-audio native perceptual backbone, jointly leveraging visual and auditory cues to support dialogue-based identity grounding, long-range event localization, and fine-grained evidence retrieval. On four standard LVU benchmarks, VideoSleuth consistently outperforms prior on-demand grounding agents and achieves competitive accuracy with the strongest dense-preprocessing baseline, while consuming at least $5.2\times$ fewer end-to-end tokens than DVD-style pipelines spend on offline preprocessing alone.

Large language models (LLMs) have shown impressive capability to understand and develop code. However, their capability to rigorously reason about and formally prove the correctness of software systems remains in question. Current benchmarks for formal proof generation focus on isolated tasks, typically algorithmic-style problems utilizing basic data structures. We curate a new benchmark suite for system-level proof generation, VeruSAGE-Bench, which consists of 849 proof tasks extracted from eight open-source Verus-verified Rust systems, covering operating systems, storage systems, memory allocators, and more. These tasks are much more complicated than those in previous benchmarks, containing on average more than 20x the formal specification (in Lines of Code) and requiring a much broader array of verification techniques. We evaluate five LLMs on VeruSAGE-Bench (o4-mini, GPT-5, Sonnet 4, Sonnet 4.5, Opus 4.5), and demonstrate that different models and agentic setups lead to vastly different success rates and costs. We also show that the best models possess impressive capability in writing correctness proofs and hence should inspire new ways of developing trustworthy software. Meanwhile, some difficult tasks in VeruSAGE-Bench stress even the best model, forcing long reasoning time (over an hour on the hardest tasks, 15.3 min per task on average), high token cost ($13.65 per task on average), and proof failures, which we hope will guide future improvement in model and agents.

Large language model-driven multi-agent systems (LLM-MAS) excel at complex tasks, yet unreliable agents remain a key bottleneck to system-level reliability. Automatic failure attribution is therefore critical, but existing approaches—such as direct prediction of agent–error pairs and agent-first failure attribution—rely on local logs of agent and miss global failures that only manifest over full interaction trajectories, such as cross-step inconsistencies and inter-agent coordination errors. Moreover, directly predicting failures induces a large combinatorial search space, hindering fine-grained attribution. To address these challenges, we propose VerifyMAS, a hypothesis verification framework for agent failure attribution. Instead of directly predicting faulty agents and error types, VerifyMAS formulates and verifies failure hypotheses against full trajectories. This verification-based approach decomposes attribution into trajectory-level error validation and fine-grained agent localization, providing an error-first attribution approach that captures global failure patterns while substantially reducing the search space. We further introduce a hypothesis-based data construction strategy grounded in a structured error taxonomy and fine-tune a specialized LLM verifier model for trajectory-level failure verification and agent attribution. Experiments on Aegis-Bench and Who&When show that VerifyMAS consistently improves diverse backbone models, including open-source Qwen and API-based GPT models, outperforming prior methods without sacrificing inference efficiency for long multi-agent trajectories.

Existing benchmarks have laid the foundation for travel planning agents by establishing API-centric paradigms. However, as the capabilities of Autonomous Agents continue to advance, their evaluation must evolve beyond simple tool execution toward handling the inherent complexities of the open web. Current benchmarks bypass core cognitive hurdles: they fail to account for information noise, ignore multi-source factual contradictions, and overlook the necessity of grounding visual perception into logical planning. We introduce VeriTrip, a verifiable benchmark designed to meet the increasing demands for agent robustness and reliability. VeriTrip shifts the evaluation focus to evidence-grounded reasoning over unstructured multimodal web corpora. It establishes a Multimodal Retrieval Base (MRB) derived from real-world sources, forcing agents to autonomously orchestrate queries across heterogeneous data. A synchronized Verifiable Knowledge Base (VKB) enables a cell-wise verification protocol that precisely quantifies factual reliability, distinguishing systematic reasoning failures from parametric hallucinations. Our evaluations across leading MLLMs reveal a critical \textit{retrieval-reasoning trade-off}: the cognitive load of autonomous retrieval significantly erodes instruction retention. VeriTrip provides the rigorous foundation necessary for the next generation of planning agents capable of operating in unconstrained, multimodal environments.

In the rapidly evolving field of Electronic Design Automation (EDA), the deployment of Large Language Models (LLMs) for Register-Transfer Level (RTL) design has emerged as a promising direction. However, silicon-grade correctness remains bottlenecked by (i) limited test coverage and reliability of simulation-centric evaluation, (ii) regressions and repair hallucinations introduced by iterative debugging, and (iii) semantic drift as intent is reinterpreted across agent handoffs. In this work, we propose Veri-Sure, a multi-agent framework that establishes a design contract to align agents' intent and uses a patching mechanism guided by static dependency slicing to perform precise, localized repairs. By integrating a multi-branch verification pipeline that combines trace-driven temporal analysis with formal verification consisting of assertion-based checking and Boolean equivalence proofs, Veri-Sure helps improve functional correctness by enriching the simulation-driven debugging loop with localized diagnostic hints. We also introduce VerilogEval-v2-EXT, extending the original benchmark with 53 more industrial-grade design tasks and stratified difficulty levels, and show that Veri-Sure achieves state-of-the-art verified-correct RTL code generation performance, surpassing standalone LLMs and prior agentic systems. Code and dataset are available at https://anonymous.4open.science/r/Veri-Sure-D861.