NeurIPS 2026 论文列表

Language models solve complex problems by articulating intermediate reasoning steps in natural language. While effective, this process is computationally bottlenecked: each reasoning step conveys only a single subword, and many are spent expressing a thought instead of carrying out computation. We propose MUX, a simple method for high-bandwidth and compact reasoning based on distillation of discrete reasoning into continuous multiplexed tokens in a latent space. Each latent token is trained to represent a weighted linear superposition (multiplexing) of a span of discrete reasoning subwords, where this superposition is lossless by construction and the span can be fully recovered (demultiplexing). We prove that simple position-dependent weightings, such as suitable geometric decay, support lossless multiplexing and prevent latent collapse. We further show that multiplexed reasoning can perform parallel exploration in problems that require search. Across 32 evaluation settings spanning four language models, MUX outperforms strong latent reasoning baselines. Our results suggest that properly chosen learning targets can make continuous latent reasoning efficient and interpretable.

Test-time scaling (TTS) improves reasoning by allocating additional inference compute, but in multimodal large language models (MLLMs), more compute is not always reliable. Longer reasoning may amplify visual misperception, induce answer drift, reinforce majority traps, or turn initially correct predictions into overthought failures. We propose MUST, a training-free framework that recasts multimodal TTS as stage-conditioned stability control. MUST separates inference into reasoning and verification stages and assigns each stage a different reliability criterion. During reasoning, Contrastive Answer-Manifold Stability (CAMS) selects answers whose supporting trajectories form compact and separable latent groups. During verification, Attention-Free Support-Chain Stability (AF-SCS) evaluates whether explicit visual evidence and subsequent reasoning form a stable latent support chain, without relying on token-to-patch attention maps. Experiments on multimodal reasoning benchmarks show that MUST improves the accuracy-compute trade-off and reduces overthinking-induced errors. These results suggest that reliable MLLM test-time scaling should adaptively control when and how extra compute is trusted, rather than simply increasing inference budget.

Recent advances in Vision Language Models (VLMs) have driven significant progress in visual reasoning. However, open-source multimodal models still lag behind proprietary systems, largely due to the lack of high-quality reasoning data. Existing datasets offer limited coverage of challenging domains such as STEM diagrams and visual puzzles, and lack consistent, long-form Chain-of-Thought (CoT) annotations essential for eliciting strong reasoning capabilities. To bridge this gap, we introduce MMFineReason, a large-scale multimodal reasoning dataset comprising 1.8M samples and 5.1B solution tokens, featuring high-quality reasoning annotations distilled from Qwen3-VL-235B-A22B-Thinking. The dataset is established via a systematic three-stage pipeline: (1) large-scale data collection and standardization, (2) CoT rationale generation, and (3) comprehensive selection based on reasoning quality and difficulty awareness. The resulting dataset spans STEM, visual puzzles, games, and complex diagrams, with each sample annotated with detailed, visually grounded reasoning traces. We fine-tune Qwen3-VL-Instruct on MMFineReason to develop MMFineReason-2B/4B/8B versions. Our models establish new state-of-the-art (SOTA) results for their size class. Notably, MMFineReason-4B surpasses Qwen3-VL-8B-Thinking, and MMFineReason-8B even outperforms Qwen3-VL-30B-A3B-Thinking while approaching Qwen3-VL-32B-Thinking, demonstrating remarkable parameter efficiency. Crucially, we uncover a "less is more" phenomenon via our difficulty-aware filtering strategy: a subset of just 7\% (123K samples) achieves performance comparable to the full dataset. Notably, we reveal a synergistic effect where reasoning-oriented data simultaneously boosts general capabilities.

Modern AI progress has been driven by ML methods that are generalizable across settings and scalable to larger regimes. As large language models demonstrate advanced capabilities in reasoning, coding, and engineering tasks, it is increasingly important to understand whether they can discover such methods rather than only apply existing ones. We introduce MLS-Bench, a benchmark for evaluating whether AI systems can invent generalizable and scalable ML methods. MLS-Bench contains 140 tasks across 12 domains, each requiring an agent to improve one targeted component of an ML system or algorithm and demonstrate that the improvement generalizes across controlled settings and scales. We find that current agents remain far from reliably surpassing human-designed methods, and that engineering-style tuning is easier for them than genuine method invention. We further study the effects of test-time scaling, adaptive compute allocation, and context provision on agents' discovery performance, together with case studies of their behavior. Our analyses suggest that the bottleneck is not only in proposing new methods, but also in the scientific insight needed to plan, validate, and scale claims about them. More search, compute, or context alone does not remove this bottleneck. We build and maintain a community platform for cumulative and comparable iteration, and release the data and code at https://mls-bench.com.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities across vision–language tasks, yet their large-scale deployment raises pressing concerns about memorized private data, outdated knowledge, and harmful content. Existing unlearning approaches for MLLMs typically adapt training-based strategies such as gradient ascent or preference optimization, but these methods are computationally expensive, irreversible, and often distort retained knowledge. In this work, we propose MLLMEraser, an input-aware, training-free framework for test-time unlearning. Our approach leverages activation steering to enable dynamic knowledge erasure without parameter updates. Specifically, we construct a multimodal erasure direction by contrasting adversarially perturbed, knowledge-recall image–text pairs with knowledge-erasure counterparts, capturing both textual and visual discrepancies. To prevent unnecessary interference, we further design an input-aware steering mechanism that adaptively determines when and how the erasure direction should be applied, preserving utility on retained knowledge while enforcing forgetting on designated content. Experiments on LLaVA-1.5 and Qwen-2.5-VL demonstrate that MLLMEraser consistently outperforms state-of-the-art MLLM unlearning baselines, achieving stronger forgetting performance with lower computational cost and minimal utility degradation.

Test-time adaptation updates a deployed model using only unlabeled target data, but standard entropy-based methods often rely on deterministic confidence and can reinforce overconfident errors under distribution shift. We propose MIRA, Mutual Information guided calibration for Reliable Test-Time Adaptation, a source-free online adaptation method that uses lightweight last-layer Gaussian perturbations to probe prediction reliability. For each target batch, MIRA reuses a single feature extraction pass and samples perturbed classifiers around the source classifier, forming an efficient local stochastic ensemble. Rather than fitting a Bayesian posterior, MIRA treats the perturbation scale as a local probing radius and selects a batch-specific scale using mutual information and prediction agreement. The resulting perturbation-sensitivity signal identifies predictions that are confident and locally stable under classifier perturbations, while suppressing brittle predictions with high stochastic disagreement. MIRA then performs adaptation with an MI-confidence weighted entropy loss, which down-weights perturbation-sensitive or low-confidence samples before they drive entropy minimization. This weighted loss reduces the influence of brittle predictions that can otherwise cause harmful adaptation under distribution shift. The method requires no source data, retraining, model ensembles, or offline posterior fitting, and can be applied as a plug-and-play modification to entropy-minimization TTA. Experiments on distribution-shift benchmarks show improvements in accuracy, calibration, and negative log-likelihood, suggesting that mutual-information-guided perturbation probing and weighted adaptation yield more robust online TTA than deterministic confidence.

Large language models (LLMs) increasingly improve their reasoning at test time via additional computation, yet most procedures treat each problem in isolation. When problems arrive sequentially, accumulating reusable experience across them can further improve performance. Existing memory-based methods either store whole-solution templates that generalize poorly to novel problems or use heuristic step-level selection that is not optimized for final-answer correctness. Learning selection policies requires large-scale training data and fixed action spaces. We propose MILES (Modular Instruction Memory with LEarnable Selection for self-improving LLM reasoning), a framework that dynamically expands step-wise memory and applies correctness-optimized memory composition under realistic test-time constraints. MILES maintains modular memory units consisting of asymmetric pairs of sub-goal embeddings and sub-instructions, each associated with a learnable selection head. This memory structure enables a coarse-to-fine retrieval mechanism: The coarse level enables memory expansion and collects supervision for training selection heads from confident samples, while the fine stage applies learned selection heads to rerank coarse-level candidates and guide reasoning for uncertain samples. MILES consistently matches or outperforms prior methods while achieving superior accuracy–efficiency tradeoffs. Extensive experiments demonstrate its effectiveness, robustness, and transferability.

A central challenge for multimodal language models (MLLMs) in real-world tasks is to abstract visual observations into structured spatial states and reason over how those states change under actions over long-horizons. Existing multimodal benchmarks often focus on static question answering, where answers can be directly extracted from images, or assume symbolic states are available and bypass the visual-to-spatial abstraction process. We present MARBLE, a diagnostic benchmark for spatially grounded multimodal reasoning and planning across three agent environments: Cube, Maze and Blox. MARBLE spans 2D and 3D spatial puzzles under geometric constraints, dynamic layouts, and spatial transformations, with difficulty systematically controlled along two axes: observation format and planning horizon. Extensive evaluations show that frontier MLLMs struggle on MARBLE, with the strongest model GPT-5.4 achieving only a 7.3\% average success rate across the three environments in the hardest setting. Pairwise ablations on the open-weight models identify visual-to-spatial abstraction as the major bottleneck: when models must infer spatial states from images instead of receiving human-annotated states, success drops from 50.7\% to 3.7\% for Gemma-4-31B and from 87.9\% to 6.5\% for Kimi-K2.6. The results indicate that models often fail to construct precise spatial states from visual observations before planning, a task that is straightfoward for human. Overall, MARBLE reveals the gap of reliable visual-to-spatial for frontier MLLMs and provides a controlled testbed for measuring progress in spatially grounded multimodal reasoning and planning.

We show that frontier AI models exhibit loyalty capture---a form of structural sycophancy in which a reporting relationship to a manager with misaligned incentives is sufficient to bias the model's recommendations, without any explicit instruction, user feedback, or adversarial prompting. In 5,757 trials across eight models from eight AI laboratories, CEO-reporting models cut investment to boost the CEO's bonus by 8.1 percentage points more when the bonus is at stake. This same effect is also predictably absent from a metric-irrelevant placebo. Delivered rationales reveal motivated reasoning: models adopt the bonus target, calculate how their recommendation achieves it, and dismiss long-term value, while concealing their recognition of the conflict from the recommendation recipient. Designed interventions (conflict disclosure, stakeholder broadening) eliminate ~60% of the effect; passive oversight does not. Taken together, the results indicate that AI models deployed within organizational hierarchies can develop and act on loyalty to their supervisor at the expense of the organization they are meant to serve.

Long-to-Short (L2S) model merging seeks to combine the strong reasoning capability of long chain-of-thought models with the concise style of base models, but existing methods that rely on linear interpolation and first-order activation statistics fail to capture how the loss landscape responds to parameter displacement, leading to degraded accuracy on hard tasks. We propose Low-Rank Hierarchical Merging (LHM), a training-free framework that replaces heuristic statistics with lightweight second-order curvature information. LHM consists of three components: (1) a low-rank Hessian approximation that efficiently estimates per-layer Hessian traces via Hutchinson's estimator on random inputs, eliminating the need for task- specific data; (2) a Layer Second-Order Sensitivity (LSOS) index that combines Hessian trace with task-vector norm to quantify per-layer fusion difficulty; and (3) a Dual-Constraint Coefficient Mapping (DCCM) that converts LSOS scores into layer-adaptive merge coefficients while jointly satisfying a fusion-error bound and a global length-reduction target. Experiments across six mathematical benchmarks on 1.5B and 7B models show that LHM outperforms state-of-the-art methods such as ACM-TIES and ORCA, achieving over 4\% accuracy gain and up to 93.1\% reduction in output length relative to the long-CoT model.

Looped language models (LoopLMs) perform iterative latent computation to refine internal representations, offering a promising alternative to explicit chain-of-thought (CoT) reasoning. However, existing reinforcement learning (RL) paradigms primarily target output tokens, creating a structural mismatch with looped architectures whose reasoning unfolds implicitly. In this work, we propose LoopRPT, a reinforcement pre-training framework tailored for LoopLMs. By reframing next-token prediction as a next-token reasoning task, LoopRPT assigns reinforcement signals directly to latent steps using an EMA teacher reference and noisy latent rollouts. This formulation enables RL to directly shape intermediate representations, compressing effective reasoning into fewer iterations. We instantiate LoopRPT on the Ouro architecture across multiple model scales. Results demonstrate that LoopRPT consistently improves per-step representation quality, achieving Pareto dominance in accuracy–computation trade-offs. Notably, significant gains on hard tokens indicate that LoopRPT enhances early-stage reasoning rather than merely encouraging premature exits. Our findings highlight reinforcement pre-training as a principled paradigm for learning efficient latent reasoning in LoopLMs.

Probabilistic logical models are a core component of neurosymbolic AI and are important in their own right for tasks that require high explainability. Unlike neural networks, logical theories that underlie the model are often handcrafted using domain expertise, making their development costly and prone to errors. While there are algorithms that learn logical theories from data, they are generally prohibitively expensive, limiting their applicability in real-world settings. Here, we introduce precision and recall for logical rules and define their composition as rule utility -- a measure of the predictive power of logical theories that can be computed in linear time. We then present SPECTRUM, a linear-time framework for learning logical theories from relational data, that mines recurrent patterns in the data, and subsequently, using our utility measure, evaluates and ranks rules derived from these patterns. Finally, we prove theoretical guarantees on the utility of the learnt logical theory. As a result, we demonstrate across various tasks that SPECTRUM scales to larger datasets, often learning more accurate logical theories on CPUs in < 1% of the runtime of SOTA neural network approaches on GPUs.

Vision-language-action models (VLA) promise human-steerable autonomous driving, but their training is bottlenecked by the scarcity of frames paired with natural-language instructions: while camera streams and expert trajectories are logged at scale, language annotations (e.g., "turn left at the intersection") remain scarce and expensive to acquire. To address this challenge, we introduce Latent Action Driving Annotations (LADA), a three-stage pipeline that transforms abundant unlabelled observation–trajectory pairs into a substrate for language-conditioned control. First, we train a latent action model with a vector-quantised bottleneck, producing a compact codebook of high-level vehicle intents. Second, a small language-annotated subset is used to train a vision-language translator to map observations and language instructions into this codebook. Third, we train a driving VLA on observation–latent-action pairs over the full unlabelled corpus. Using fewer than 5% of language annotations and without leveraging any auxiliary chain-of-thought reasoning or visual question answering streams, LADA achieves a Driving Score of 87.98 and a Success Rate of 70.46% on the closed-loop Bench2Drive benchmark, matching or surpassing fully supervised baselines.

Understanding human behavior while interacting with the surrounding world is crucial for many applications of embodied AI. First-person videos are particularly informative for this problem, as they well capture how activities reshape the scene over time. However, existing approaches often rely on implicit visual or language-aligned representations, disregarding structured reasoning over the scene dynamic. In this paper we argue that explicit, compositional and editable representations of human-environment interactions can play a crucial role for rich grounded activity understanding. To this end, we introduce SG-Ego, a large scale annotation set extending Ego4D with spatio-temporal scene graphs, where relations triplets are consolidated over time into explicit time-evolving descriptions of the scene state. To effectively reason over this representation, we propose GLEN, a graph-based model that operates over scene graph sequences to both align them with textual actions and model their temporal evolution. In addition, we formulate the activity-driven graph-edit forecasting (A-GEF) problem, a novel task that casts scene dynamics as a sequence of structured transformations conditioned on ongoing actions, enabling explicit reasoning about how scenes change over time. We demonstrate the effectiveness of our approach across multiple downstream tasks, spanning retrieval benchmarks as EgoMCQ and the more complex compositional EgoCVR, as well as long-horizon reasoning benchmarks as EXPLORE-Bench and the newly introduced A-GEF. GLEN achieves competitive or state-of-the-art performance compared to raw video baselines and it excels in reasoning settings, typically addressed only with MLLMs, while enabling controllable and structured predictions of scene dynamics driven by human activities. We believe our results establish spatio-temporal scene graphs, together with models that reason over them, as strong compositional and interpretable representations for egocentric video understanding and potentially beyond. Code will be released upon acceptance.

Embodied LLMs endow robots with high-level task reasoning, but they cannot reflect on what went wrong or why, turning deployment into a sequence of independent trials where mistakes repeat rather than accumulate into experience. Drawing upon human reflective practitioners, we introduce Reflective Test-Time Planning, which integrates two modes of reflection: \textit{reflection-in-action}, where the agent uses test-time scaling to generate and score multiple candidate actions using internal reflections before execution; and \textit{reflection-on-action}, which uses test-time training to update both its internal reflection model and its action policy based on external reflections after execution. We also include retrospective reflection, allowing the agent to re-evaluate earlier decisions and perform model updates with hindsight for proper long-horizon credit assignment. Experiments on our newly-designed Long-Horizon Household benchmark and MuJoCo Cupboard Fitting benchmark show significant gains over baseline models, with zero-shot generalization to photorealistic HM3D environments and real-robot experiments on a Franka Panda arm. Ablations confirm that reflection-in-action and reflection-on-action are mutually dependent, and that retrospective reflection achieves better credit assignment than step-wise external feedback at lower computational overhead. Qualitative analyses further highlight behavioral correction through reflection.

Reinforcement learning from verifiable rewards (RLVR) produces strong reasoning models, yet these models can fail catastrophically when the conditioning context is fallible (e.g., corrupted chain-of-thought, misleading partial solutions, or mild input perturbations), because standard RLVR optimizes final-answer correctness only under clean conditioning. We introduce GASP (Guided Adversarial Self-Play), a robustification method that explicitly trains error detection and repair capabilities using only outcome verification. Without human labels or external teachers, GASP instantiates an adversarial self-play game within a single model: a polluter learns to induce failure through locally coherent corruptions, while an agent learns to diagnose and recover under the same corrupted conditioning. To address the scarcity of successful recoveries early in training, we propose in-distribution repair guidance, an auxiliary imitation objective on self-generated repairs that increases recovery probability while preserving previously acquired capabilities. Across four open-weight models (1.5B–8B), GASP converts strong-but-brittle reasoners into substantially more robust ones that better withstand misleading and perturbed context, while often improving clean accuracy. Further analysis shows that adversarial corruptions create an effective curriculum, and that in-distribution guidance enables rapid recovery learning with minimal representational drift.

Modern reasoning models depend on reasoning data, today sourced from human annotations or distilled from stronger LLMs. However, a rich and largely untapped source of supervision lies in expert systems (e.g., game engines, classical planners, theorem provers), which routinely produce near-optimal actions across diverse domains. But these experts are silent: they commit to an action without writing down the chain of thought (CoT) behind it. Recovering that CoT as natural-language reasoning would distill expert knowledge into a student that generalizes beyond the demonstrated actions. We treat it as a latent variable and study how to recover it from the action alone. Our approach, LeAct (**Le**arning to reason from **Act**ions), optimizes this latent variable: the student samples candidate CoTs for each expert action, and we retain those that measurably improve its own probability of recovering the action. Across imperfect-information games at multiple scales and a simulated robotics benchmark, LeAct reaches the solver's numerical floor on small enumerable games. At larger scale, it is $5\times$ closer to the solver than the strongest expert-iteration baseline. At Flop Hold'em ($\sim 10^9$ infosets), LeAct wins head-to-head by $+60$ mbb/g, and on the robotics probe it is the only training recipe that improves on direct imitation. We present a principled framework and the result: expert systems become a categorically new source of reasoning teachers for foundation models.

While joint audio-visual understanding is fundamental to advancing machine cognition, current multimodal large language models (MLLMs) still struggle with complex cross-modal reasoning. Existing text-based chain-of-thought (CoT) compresses rich multi-modal features into discrete text, incurring information loss and inducing a language-bound phenomenon that diminishes attention to audio-visual signals. In contrast, a continuous latent space inherently preserves dense representations, serving as an ideal carrier for audio-visual information. Motivated by this, we propose $\textbf{LatentOmni}$, a novel cross-modal reasoning framework. By introducing a feature-level supervision mechanism to directly reconstruct raw sensory inputs within the latent space, LatentOmni leverages native latent features to bridge audio-visual modalities and text, ensuring sustained attention on original audio-visual inputs throughout reasoning. Furthermore, to maintain temporal consistency across modalities in latent space, we design Omni-Sync Position Embedding (OSPE), which generalizes multimodal rotary position encodings to drive audio-visual synchrony. To supervise this reasoning process, we construct LatentOmni-Instruct-35K, a dataset interleaving text with audio-visual segments that serve as dense evidence for latent reconstruction. Comprehensive evaluation across multiple audio-visual reasoning benchmarks demonstrates that LatentOmni substantially outperforms strong explicit-CoT baselines, validating latent space joint reasoning as a promising path toward genuine omnimodal understanding.

Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model’s residual stream. Despite their empirical success, these methods lack a principled account of the latent-space transformation they induce and why it suppresses refusal. In this work, we recast refusal suppression as a latent-space evasion attack against linear probes trained to separate refused from answered prompts. Under this view, prior work’s difference-in-means direction naturally defines such a probe, and its ablation is exactly a projection onto its decision boundary, i.e., a minimum-confidence evasion attack. This perspective not only explains the empirical success of prior work but also admits a key limitation: evasion stops at the decision boundary, motivating the need to push representations further into the compliant region, i.e., where the model answers. We leverage this by proposing a Controlled Latent-space Evasion attack that projects representations past the boundary with an optimized confidence. We achieve state-of-the-art attack success rate across 15 instruction-tuned, multimodal, and reasoning models, outperforming existing refusal-ablation baselines and specialized jailbreak attacks.