NeurIPS 2026 Papers

As LLM agents advance rapidly in autonomy, the risk of hidden objectives that bypass safety constraints becomes increasingly plausible. Chain-of-thought (CoT) monitoring offers a key control mechanism by reading the agent's reasoning and flagging misalignment before it materializes into action, yet its reliability depends on the realism of the red-team it is stress-tested against. Existing red-teaming efforts apply uniform attack templates that treat the side task as an exogenous insertion, yielding abrupt anomalies any reasonable monitor can flag while leaving the subtler, contextually grounded evasions of a capable agent largely unprobed. We therefore ask: How robust is CoT monitoring against agents that strategically conceal misalignment? To answer this, we introduce TraceMRT (Trace-level Monitor Red-Teaming), an automated red-teaming framework that reframes the side task as a coherent sub-goal of the primary task and searches trace-shaping strategies whose induced trajectories complete the misaligned objective while reading as natural execution. Through black-box attacks on state-of-the-art monitors (e.g., GPT-5.2, and Claude-4.5), we reveal substantial vulnerabilities even in top models. To better counter these vulnerabilities, we further improve monitoring scaffolding with Bayesian Trajectory Monitoring, integrating global context and local evidence to better detect covert misalignment. Empirically, the method significantly consolidates the monitor with 0.91 AUC and 0.74 TPR@FPR=0.05.

Streaming video understanding requires models not only to process temporally incoming frames, but also to anticipate user intention for realistic applications such as Augmented Reality (AR) glasses. While prior streaming benchmarks evaluate temporal reasoning, none measure whether Multimodal Large Language Models (MLLMs) can interpret or leverage human gaze signals within a streaming setting. To fill this gap, we introduce StreamGaze, the first benchmark designed to evaluate how effectively MLLMs utilize gaze for temporal and proactive reasoning in streaming videos. StreamGaze introduces gaze-guided past, present, and proactive tasks that comprehensively assess streaming video understanding. These tasks evaluate whether models can use real-time gaze signals to follow shifting attention and infer user intentions based only on past and currently observed frames. To build StreamGaze, we develop a gaze–video Question Answering (QA) generation pipeline that aligns egocentric videos with raw gaze trajectories through fixation extraction, region-specific visual prompting, and scanpath construction. This pipeline produces spatio-temporally grounded QA pairs that reflect human perceptual dynamics. Across all StreamGaze tasks, we observe substantial performance gaps between state-of-the-art MLLMs and human performance, highlighting key limitations in gaze-based temporal reasoning, intention modeling, and proactive prediction. We further provide detailed analyses of gaze prompting strategies, reasoning behaviors, and task-specific failure modes, offering insights into current limitations and directions for future research. All data and code are publicly available to support research in gaze-guided streaming video understanding.

While large reasoning models (LRMs) can solve complex tasks via generating extended Chain-of-Thought traces (Long-CoT), longer traces often increase confidence without increasing correctness. In this work, we formalize such failure via a \emph{self-conditioned evidence} model: generated tokens provide information about the model's current working hypothesis rather than the ground-truth answer. This yields a simple diagnostic principle: entropy, fluency, and answer stability can certify within-trajectory self-consistency while remaining causally disconnected from correctness. Building on these findings, we propose \textbf{Faithful Pandora Inference (FPI)}, a minimal training-free controller that keeps the base LRM frozen and fits only a small calibration map on held-out data. At each checkpoint, FPI estimates calibrated correctness $r_t$ and the marginal evidence gain $m_t$ of continuing the current trajectory. It stops only when $r_t$ reaches a user-specified reliability target; if $r_t$ remains insufficient and $m_t$ falls below cost, it reallocates the remaining budget to a diversified restart. Across six reasoning benchmarks and multiple DeepSeek-R1 distilled models, FPI improves the accuracy-efficiency frontier, reduces expected calibration error, and improves selective accuracy on high-confidence outputs. Ablations on confidently wrong trajectories show that calibrated stopping and marginal-gain restarts address different failure modes and that the gains are not explained by additional samples alone.

Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model's on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome–process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 56.3% on average while improving accuracy by 4.2% relative to the base model.

Flexible switching between reasoning trajectories (i.e., thoughts switching) has significantly enhanced the reasoning capabilities of Large Reasoning Models (LRMs). However, existing models often switch excessively yet fail to sustain promising reasoning thoughts---a phenomenon termed ''under-thinking''. While recent efforts suppress switching to mitigate this, such over-correction may discard valuable trajectories. To address this challenge, we propose Steady Thought (ST), a novel thought-level preference optimization framework. ST formalizes under-thinking as a preference issue at switching points, forcing single-path continuations from identified points to yield high-quality reasoning. Then, ST performs thought-level preference optimization by treating the newly generated response as preferred and the original one as dis-preferred. Experiments across multiple models and datasets show that ST effectively reduces token consumption while maintaining or even improving accuracy. It reduces output length by up to 63.8% while improving accuracy by up to 11.4%. Further analysis suggests that ST helps models acquire a generalizable ability to reduce redundant switching across domains and languages.

Test-time compute has emerged as a major approach to improving the capabilities of Large Language Models (LLMs). However, existing test-time reasoning paradigms rely heavily on externally imposed control, either through fixed reasoning programs or through costly expansion in constrained search spaces, limiting both generalization and efficiency. We propose State of Thought (SoT), a new reasoning paradigm that enables endogenous reasoning in LLMs, with the model's internal reasoning state governing how reasoning unfolds. Concretely, SoT extracts a compact dynamics-geometric state from the model's internal information transfer and selectively activates historical reasoning support useful under the current reasoning state, framing reasoning as a state-conditioned process over evidence rather than an externally prescribed token chain. Across quantitative (1.29x), general (1.62x), symbolic and code (1.72x), long-context (2.63x), and multimodal (1.08x) reasoning tasks on 4 models with 20 datasets, SoT consistently improves task accuracy while reducing tokens by 69.0% and latency by 48.7%, supporting endogenous state-driven reasoning as a more generalizable and efficient alternative.

We study the accumulation of test errors in language models performing multi-step reasoning. Although longer reasoning improves model capabilities by scaling test-time computation, test error compounds in autoregressive generation and can grow substantially across steps. The key factors determining the stability of such reasoning remain unclear. To address this question, we analyze the test error of multi-step reasoning under supervised fine-tuning. We first present a test error bound governed by the product of *input Jacobian spectral norms* across generation steps. This product, summarized as an error amplification factor, scales exponentially with steps and controls the reasoning stability. Then, we explicitly analyze this factor in transformers trained to predict linear and quadratic function weights. We prove that the transformer converges to a solution where this amplification factor decays, yielding near-zero test loss even over longer steps. Based on the analysis, we propose a training method to suppress the error amplification factor by combining (i) chain-of-thought length compression that reduces the reasoning steps, and (ii) quantization-aware training that provably regularizes the input Jacobian norms. We validate our method by fine-tuning language models on reasoning tasks that require executing graph algorithms and tracking variable states through logical operations. Across seven tasks, our method improves over baselines by 3.5\% on average, and by 8.2\% in length generalization evaluations. We further validate that it reduces the error amplification factor in fine-tuned models.

Multi-agent orchestration over large language models is increasingly used to improve reasoning accuracy, especially on problems that benefit from critique, verification, specialization, or consensus, but these gains often come with inference costs that are not explicitly controlled. Many systems spend similar compute on straightforward queries and genuinely ambiguous or expert-level ones. Existing methods typically follow fixed collaboration protocols, compile task-level workflows offline, route queries among models or collaboration modes, or use learned controllers whose state does not explicitly identify which reasoning defects remain unresolved. As a result, they do not directly decide during inference whether the current multi-agent trajectory still justifies another agent call. We introduce Defect-Aware Residual Coverage (DARC), a training-free inference-time controller for adaptive multi-agent reasoning that represents the current trajectory using four residual defect dimensions: answer uncertainty, claim-level contradiction, verification failure, and aspect under-coverage. DARC selects the next agent by maximizing cost-normalized submodular marginal coverage over the remaining residuals. Each candidate agent receives a role-induced capability profile from frozen embeddings of its natural-language role description, enabling DARC to operate without task-specific supervision, learned routing parameters, or fine-tuning. Across five reasoning and multimodal suites, MMLU-Pro, GPQA-Diamond, LiveBench, MMMU-Pro, and HLE, DARC achieves the strongest accuracy under the main shared heterogeneous six-agent pool while using substantially fewer tokens and agent calls than recent workflow, adaptive orchestration, and cost-aware routing baselines, with additional large-pool experiments showing consistent scaling behavior. It also outperforms the single-call frontier reasoning models tested, showing that residual-aware selective invocation can improve both answer quality and inference efficiency.

We prove that the speedup of exact speculative decoding is governed by a per-token rate, not by the path-level prefix-overlap capacity that folklore suggests, and that the gap between the two can be unbounded. The two candidates are the per-token Leviathan rate σ_L(P, Q) and the path-level coupling capacity Ω_L(P, Q) = Σ_{t=1}^L ov(P_{≤t}, Q_{≤t}). We show σ_L ≤ Ω_L always, the inequality is strict already on i.i.d. products, and there exists an explicit fixed-precision binary trajectory-exact rejection-style protocol on L = 2 with E[τ] = 19/16 > 9/8 = σ_L, ruling out σ_L as a universal upper bound across all trajectory-exact rounds. Inside the deployable admissible Leviathan-style clipped-ratio class, σ_L is the tight rate, and the speculation decoding complexity satisfies SDC_{B,L}^Q(M; u, T) = Θ(T / (σ_{L,B}(P_M^{u,T}, Q) + 1)), achieved by iterated Leviathan and matched up to constants by Wald. On the path side, Ω_L exhibits a sharp two-sided dichotomy in the prefix-KL profile η_t: Σ_t √(η_t / 2) = o(L) forces Ω_L = L − o(L), while geometrically bounded overlap o_t ≤ ρ^t forces Ω_L ≤ 1/(1−ρ) uniformly in L, with no unconditional KL converse. Two explicit fixed-precision Transformer families of size O(log L) realize the two extremes under the same three-element proposal family Q = {Q_flat, Q_const_0, Q_const_1}: a constant-readout family F⁺ has SDC ≤ T / ((1−ε)L + 1), while an alternating-parity family F⁻ has SDC ≥ T/5, an unconditional Θ(L) architecture separation. Speculative decoding is not path coupling; it is per-token clipped-ratio acceptance, and the architecture of the target controls the rate.

Multimodal large language models (MLLMs) excel at visual interpretation but fail on spatial reasoning tasks that humans solve reliably. Existing benchmarks evaluate these models as black boxes, limiting their ability to identify the underlying causes of lower performance: when a model fails a spatial reasoning task, it remains difficult to ascertain whether the hurdle is perceptual, such as recognizing object boundaries, or cognitive, such as reasoning about occlusion to infer hidden geometry. We introduce Spatial-IQ, a hierarchical diagnostic framework that decomposes object counting in stacked 3D structures into 9 perceptual and cognitive sub-tasks organized by the developmental stages of human spatial cognition, with mental rotation as an additional target probe. Using NVIDIA Isaac Sim, we procedurally generated a diverse dataset of roughly 80,000 stacked 3D structures with per-task ground truth. We evaluate models across three output formats (free-response text, multiple-choice images, and image editing) alongside a human baseline. The Spatial-IQ framework shows that top-performing models often succeed at the target task (object counting) without succeeding on the lower-level sub-tasks intended to support it, and that models differ in how much of these hierarchical chains they preserve, often revealing shortcut behavior that raw target-task accuracy alone would obscure. Finally, we demonstrate that training models with chain-of-thought (CoT) supervision over our hierarchical sub-tasks, combined with reinforcement learning with verifiable rewards, significantly improves both spatial consistency across sub-tasks and target-task accuracy, supporting the value of the proposed decomposition as both a diagnostic tool and a training signal.

LLM chain-of-thought reasoning improves with additional compute, but standard 2 inference scaling [Wei et al., 2022, Wang et al., 2022, Snell et al., 2024, Brown 3 et al., 2024] treats all steps equally. We ask: which step in a reasoning chain benefits 4 most from extra computation? We make three layered contributions, each scoped 5 explicitly. (I) Measurement. We define expansion utility Uk as the marginal 6 gain in correctness probability from resampling from step k onward, measured via 7 a step-oracle sweep. Across nine model families (8B–123B), seven benchmarks 8 (math, science, code), and 6,500 chains, every one of the 41 healthy (model, bench) 9 cells shows positive oracle gap at 10% step budget (95% Wilson CI [0.914, 1.000]; 10 mean +22.0 pp, range +1 to +48 pp). (II) Structure. Per-step utility is sparse 11 (97.6% of cells: Gini ≥ 0.80); Gini and mean oracle gap are tightly anti-correlated (Pearson r = −0.913, p = 9.5 × 10−17 12 ). A resample-stability simulation and 13 a segmentation-rule sensitivity check (Appendix J) bound the headline against 14 noise and rule-choice confounders. (III) Routing proof-of-concept. A cross15 model router trained with a within-example listwise loss [Cao et al., 2007] and a 16 learned depth prior, combined at inference with a bench-conditional depth mask 17 and a confidence gate, recovers 5-seed-averaged gated GR@10% = +0.149 (95% 18 bootstrap CI [+0.085, +0.217]) of the cross-model oracle gap at matched step19 budget, with no architectural change to the underlying model. We frame this as 20 proof the framework is actionable cross-model, not as deployment-ready scaling; 21 matched-token comparison to self-consistency / best-of-N requires a separate 22 generation campaign and is left to follow-up (§7). Methodological observation. 23 Within-example pointwise top-k BCE under cross-model pooling produces a high24 AUC discriminator (AUC@10 = 0.70) that selects the wrong steps (gated GR 25 = −0.15); the listwise loss is shift-invariant to per-(model, bench) utility scale and 26 removes this inversion.

Reinforcement learning with verifiable rewards (RLVR) has greatly advanced large reasoning models (LRMs), but it requires timely training on a huge fully-annotated dataset. To this end, data-efficient RLVR methods have been widely studied from two perspectives: (i) data selection methods identify a small subset of "golden" samples that yield near-full-data performance, but they rely on a pre-existing pool of labeled data. (ii) unsupervised RLVR methods train the model using its own internal supervision signals on large-scale unlabeled data, yet they exhibit suboptimal performance. Accordingly, we investigate the "*pick in the dark*" setup for RLVR, which aims to select, without prior supervision, unlabeled samples that are most beneficial for training and worthy of annotation. Through systematic analysis, we demonstrate that smart picks hinge on a well-calibrated uncertainty estimator to enable strategic partitioning of data for adaptive training regimes. Building on this insight, we propose **PivotTrace**, a three-way data triage framework that leverages attention dynamics to trace *metacognitive pivots* during reasoning. By precisely quantifying uncertainty through pivot density, PivotTrace achieves automated data routing to synergistically maximize both annotation and training efficiency. Empirically, PivotTrace surpasses the fully supervised LRM with only **29.3%** annotated samples and **2.75**$\times$ faster convergence.

Plot-to-code generation has been greatly advanced by recent multimodal large language models (MLLMs). However, existing studies mainly focus on one-pass generation or iterative sampling guided by verifier-based filtering, which does not align with the typical human programming workflow of generate first, then iteratively error-location and fix. This naturally motivates a test-time refinement paradigm for plot-to-code generation.In this work, we present the first study on plot-to-code generation via iterative test-time refinement. We propose a Visual-Code Diagnostics framework that enables refinement through visual feedback from source code. Specifically, we first categorize generation errors into four coarse-grained types, and then train lightweight discriminators to recognize and localize these errors based on the rendered visual outputs and corresponding code. Unlike traditional compiler- or debugger-based verification, our approach performs iterative refinement guided by one diagnostics system, enabling more effective semantic and visual correction during test time. Extensive experiments demonstrate that our method consistently improves generation quality across diverse generators and significantly outperforms strong test-time scaling baselines.

Self-improvement at scale has been a longstanding goal for reasoning models, and there are two natural places to do it: at test time, through verification-refinement (V-R) loops; and at training time, through self-training methods. Both are gated by the same bottleneck: the verifier. V-R loops stall when verifier scores inflate over rounds while accuracy stagnates, and when feedback is too generic to act on; self-training fails similarly when bad self-generated data are added to training. Better verifiers would unlock both, but the capability we want to train, i.e., catching errors the generator cannot detect in its own work, has neither finetuning data nor a verifiable reward. To address this challenge, we propose *self-trained verification* (STV). Our key observation is that, while a model cannot catch these errors alone, it can when shown the reference solution. We turn this asymmetry into a supervision target and train the verifier to imitate a more informed version of itself. At test time, STV substantially improves V-R loops on hard problems, while standard alternatives (e.g., SFT, RL on verifier scores, and even meta-verifiers) do not. STV roughly doubles accuracy on hard DAPO math and lifts it 14x on the hardest SciKnowEval problems (1.5% to 21%). At training time, starting from an RL-converged generator, putting the STV verifier in the loop yields a further 33% relative gain in test-time pass@1. More notably, the generator's standalone pass@1, with no verifier at inference time, climbs 30% relative past where continued RL had converged. Hence, the next frontier in reasoning may lie in how we train verifiers.

Multimodal LLMs often produce fluent yet unreliable reasoning, exhibiting weak step-to-step coherence and insufficient visual grounding, largely because existing alignment approaches supervise only the final answer while ignoring the reliability of the intermediate reasoning process. We introduce SR-MCR, a lightweight and label-free framework that aligns reasoning by exploiting intrinsic process signals derived directly from model outputs. Five self-referential cues—semantic alignment, lexical fidelity, non-redundancy, visual grounding, and step consistency—are integrated into a normalized, reliability-weighted reward that provides fine-grained process-level guidance. A critic-free GRPO objective, enhanced with a confidence-aware cooling mechanism, further stabilizes training and suppresses trivial or overly confident generations. Built on Qwen2.5-VL, SR-MCR improves both answer accuracy and reasoning coherence across a broad set of visual benchmarks; among open-source models of comparable size, SR-MCR-7B achieves state-of-the-art performance with an average accuracy of 81.4%. Ablation studies confirm the independent contributions of each reward term and the cooling module.

Language models routinely revise initial answers through intermediate edits, yet current alignment objectives restrict credit assignment to complete outputs or entire rollouts. To formalize the local geometry of improvement, we introduce Lifted State Policy Optimization (LSPO), which augments each textual answer with a continuous auxiliary coordinate to form a lifted state that encodes refinement context beyond surface text. Training rewards transitions that descend an energy landscape, subject to edit and step penalties, and internalizes the resulting answers into the generative model. Across the Qwen3.5 family on mathematics, science, code, logic, and broad reasoning benchmarks, LSPO improves aggregate accuracy while bypassing the latency of explicit revision loops. Component ablations, together with evaluations of internalization and energy ordering on unseen tasks, confirm that this gain originates from the lifted representation and from credit assigned at the transition level.

Today, verifier-guided best-of-$N$ inference is a common way to spend test-time compute in language modeling, code generation, and alignment: generate $N$ candidates, score them with a verifier or reward model, and deploy the highest-scoring output. This paper argues that the relevant reliability object is not average verifier accuracy but \emph{selected-tail reliability}, namely the behavior of the output chosen after optimizing over $N$ noisy public scores; we derive exact selected-rank exposure laws for best-of-$N$ selection and finite-verifier resource laws showing that selected public optimism scales at the $\sqrt{\log N/m}$ regime under $m$ units of verifier evidence. We then separate public optimism from hidden utility harm, showing that public-score inflation alone is not a harm claim and that harm or uncertifiability requires public-hidden mismatch, tail dominance, count imbalance, or an unresolved selected-tail regime. Building on these results, we introduce Tail-Certified Compute Caps, a conservative procedure that certifies candidate budgets or refuses certification using selected-tail intervals, verifier-resource penalties, mismatch envelopes, and feasibility gates. In exact-verifier/public-hidden experiments, the theory predicts selected false-positive exposure and supports resource-aware cap decisions over an all-attempted denominator, while non-code candidate tables with pre-existing public human annotations serve as secondary scoped $K\ge16$ consistency checks under proxy/model-prior/verifier-style scores across Arena55K and Stanford SHP. Overall, the paper provides a reliability framework for verifier-guided inference-time scaling: candidate budgets should be justified by evidence about the selected tail that is actually deployed, not only by average verifier validation.

We present Seirênes, a self-play RL framework that transforms contextual interference from a failure mode of LLM reasoning into an internal training signal for co-evolving more resilient reasoners. While RL with verifiable rewards has significantly advanced reasoning capabilities, models can still exhibit fragility when encountering non-idealized contexts: scenarios characterized by superfluous information, tangential instructions, or incidental correlations that differ from the clean distributions typical of standard benchmarks. Seirênes harnesses this vulnerability through a parameter-shared and adversarial self-play loop. Within this framework, a single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots, and solve problems by discerning the essential task from these perturbations to recover the core underlying logic. By pitting these competing objectives against each other, Seirênes compels the model to move beyond superficial pattern matching and anchors its capabilities in robust underlying reasoning. This continuous interaction sustains an informative co-evolutionary curriculum as the model improves. Across seven mathematical reasoning benchmarks and model scales from 4B to 30B, Seirênes achieves average gains of +10.2, +9.1, and +7.2 points. Besides, distracting contexts produced by the 4B Seirênes model reduce the accuracy of top-tier closed-source models (GPT and Gemini) by roughly 4--5 points, revealing Seirênes' general ability to uncover reasoning models' blind spots. Our code will be available.

3D multimodal language models are becoming a foundation for embodied agents, indoor robotics, and human–AI interaction, but their answers and explanations often remain weakly tied to the specific objects and spatial relations observed in a 3D scene. This paper aims to make 3D reasoning more faithful by reducing language-prior rationalization and forcing reasoning chains to depend on verifiable scene evidence. We propose Anti-Rationalization Chain-of-Thought (AR-CoT), a plug-in training and decoding framework that represents each reasoning step with an explicit evidence pointer and scores candidate answer-chain pairs by their scene-evidence gain. AR-CoT contrasts scene-conditioned rationales with a frozen text-only prior, verifies local object and relation claims, and uses decoy-twin scene pairs to encourage reasoning chains to change when the relevant spatial evidence changes. Experiments on standard 3D QA and grounding benchmarks, as well as MSQA, Beacon3D, and shortcut-sensitive stress tests, show that AR-CoT consistently improves multiple 3D MLLM backbones while strengthening chain grounding, decoy-twin contrast, and scene-swap sensitivity. These results suggest that anti-rationalization offers a practical and verifiable route toward more accurate, interpretable, and scene-faithful 3D multimodal reasoning.