NeurIPS 2026 论文列表

Large language models (LLMs) have shown promise in medical question answering and clinical reasoning, yet their improvement remains constrained by static parametric knowledge and costly expert supervision. Self-evolving agents offer a promising alternative by enabling models to improve through iterative task generation and problem-solving. However, most existing self-evolving methods are designed for easily verifiable domains such as mathematics and coding, where solutions can be checked by exact answers or executable programs. Medical reasoning is fundamentally different: it is open-ended, knowledge-intensive, and often only partially verifiable. We present \textbf{MedZERO}, a self-evolving framework for open-ended medical reasoning. MedZERO couples an \textit{Examiner} that generates frontier medical question-option pairs with a \textit{Reasoner} that solves them through evidence-grounded multi-turn reasoning with external knowledge tools. To support reliable, continual improvement, MedZERO introduces controlled knowledge accumulation, which maintains temporary exploratory knowledge and curated persistent knowledge in reasoning. We evaluate MedZERO on five public medical reasoning benchmarks using 4B- and 8B-scale base models under open-ended evaluation. Across all settings, MedZERO consistently outperforms the underlying base models and prior self-evolving baselines, achieving up to 13.7 average accuracy-point gains over the next-best self-evolving baseline.

Language models are increasingly trained to "reason" before answering users' queries, outputting hundreds of intermediate tokens before a final answer. While reasoning is designed with direct model capabilities in mind, we argue that reasoning language models (RLMs) should also be assessed by interaction effects between their reasoning traces and weaker models. In contexts like safety monitoring and distillation, the performance of weak models increasingly depends on legibility---how thoroughly and precisely a reasoning trace externalizes an RLM's actions. Existing efficiency-based metrics for legibility fail to capture "thoroughness", instead focusing on conciseness. Thus, we introduce *transfer utility*, a method for measuring trace legibility derived from interactions where weak models finish tasks given incomplete traces. Evaluating 85k traces from 12 RLMs across three tasks, we find that reasoning traces generated by the highest-performing and the most efficient models rank the *lowest* for transfer utility. We also find evidence that transfer utility predicts the ability of weak monitors to verify procedurally dense traces for math or logical problem-solving, though the effect moderates when verifying fact-driven traces. Finally, we show that open-source reward models decouple from transfer utility signals on correctness. Together, these findings surface status quo trade-offs in weak-to-strong legibility, motivating an interaction-first framework for future applications.

Designing tool-integrated reasoning agent systems for complex chemical tasks remains a fundamental challenge. Although recent chemical agents have demonstrated promising performance through the integration of external tools, they are still built on the autoregressive paradigm, which is inherently constrained by causal attention and left-to-right generation. As a result, they often struggle with bidirectional molecular understanding, global coordinated tool planning and multi-source evidence integration. To address this challenge, we introduce ChemDiffAgent, the first masked diffusion language agent for tool-integrated chemical reasoning. It reformulates chemical reasoning as iterative denoising over interleaved reasoning and tool-use trajectories, enabling tool-use decisions to be jointly determined under full-trajectory context and thereby better aligning the agent's decision boundary with its knowledge boundary. We further analyze how this paradigm benefits agentic reasoning through harder infilling subproblems during training and flexible arbitrary-order decoding during inference. Then we develop a two-stage post-training framework, comprising agentic supervised fine-tuning and variance-reduced preference optimization, to enhance tool-use and reasoning capabilities. We also construct a comprehensive benchmark covering six chemistry tasks across both single-turn and multi-turn tool-use scenarios. Under matched budgets and evaluation, ChemDiffAgent consistently outperforms autoregressive agents, achieving an overall improvement of approximately 20\%. Notably, our 8B model achieves performance comparable to, in some cases surpassing, significantly larger frontier models, such as GPT 5.5 and Claude Opus 4.7.

Vision-language models (VLMs) can achieve strong image-level visual question-answer (QA) performance while relying on shortcuts rather than grounded visual understanding. This is especially concerning in medical imaging, where clinically meaningful interpretation often requires spatial reasoning, localization, and anatomical understanding. Yet in mammography, no benchmark directly assesses whether VLMs can localize findings or reason spatially. To address this, we introduce MammoGPS, a benchmark for evaluating visual grounding, perception, and spatial reasoning in mammography. MammoGPS comprises 56,638 QA pairs across 13 clinical tasks for anatomical and finding grounding, finding-type, quadrant prediction, and depth determination. MammoGPS also includes 94,119 QA pairs across 20 controlled diagnostic tasks designed to probe model failures and spatial biases: we design controlled difficulty ladders to isolate the effects of semantic cues, visual cues, clinical terminology, and background context, as well as spatial-bias tasks to test whether model performance depends on finding location. We evaluate 12 VLMs spanning proprietary, general-purpose, and medical-domain models. Our analysis shows that current systems struggle to localize suspicious findings, make limited use of spatial semantic cues, rely on priors and shortcut heuristics for categorical and grounding tasks, struggle with mammography-specific spatial terminology, and exhibit position-dependent biases. Together, our results underscore the need for more rigorous and fine-grained mammography VLM evaluations.

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.

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.

Mobile agents are increasingly expected to operate everyday applications from screenshots and language goals, where reliable control requires reasoning over screen affordances, multi-step navigation, and future state changes. Yet many agents externalize this computation as long textual thoughts, making interaction slower, supervision more costly, and deployment less efficient. We introduce MIRAGE (Mobile agents with Implicit Reasoning And Generative world modEls), a framework that learns continuous latent reasoning representations from visible textual thoughts. MIRAGE introduces an efficient latent-space learning procedure that transfers explicit reasoning into compact hidden states, allowing the agent to reason internally without decoding long rationales. It further brings a world-model perspective into mobile-agent training: the model’s latent reasoning vectors are aligned with future screenshots, encouraging the agent to predict upcoming interface states in latent space before executing an action. This makes the hidden computation not only a compressed thought trace, but also a forward-looking representation of how the environment may change. At inference time, MIRAGE reasons in continuous latent space, reducing token generation while improving execution efficiency. On AndroidWorld, MIRAGE matches explicit-CoT SFT in the 4B ablation under a 3–5× lower decoded-token budget, and improves a comparable instruction-tuned baseline by 10.2 points; on AndroidControl, it improves action grounding with over 75% fewer generated tokens.

While Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced multimodal reasoning, existing frameworks suffer from a critical “perception–reasoning gap.” Due to an over-reliance on seemingly relevant visual tokens, even minor perceptual perturbations can propagate through the reasoning process, leading to compounding hallucinations. To address this issue, we propose MIRA, a training framework that improves robustness to erroneous visual contexts while enhancing logical reasoning ability. By injecting filtered, deceptive visual contexts during training, we construct challenging “hard traps” that stress-test the model’s robustness to misleading information. Furthermore, we introduce a group-wise reflection-triggering mechanism that enables the model to autonomously detect and rectify inconsistencies between visual cues and logical constraints, which is ultimately internalized as an intrinsic policy behavior. Importantly, MIRA does not rely on human annotations. Experimental results across multiple benchmarks show that MIRA improves the base model by 7% and demonstrates superior robustness against deceptive contexts. These findings highlight that active error recovery is essential for reliable multimodal reasoning. Comprehensive ablation studies and analyses further provide insights into why MIRA is effective.

Large language model agents are increasingly deployed through the Model Context Protocol (MCP), which allows them to discover and invoke tools across multiple servers at runtime. Existing MCP benchmarks primarily measure task completion capability but provide limited insight into why agents fail, reporting aggregate success rates without attributing failures to specific reasoning, execution, or memory breakdowns. We introduce MCPHallu, a benchmark for diagnosing hallucination-prone agent failures in MCP environments along three axes: reasoning, execution, and memory. MCPHallu contains 358 tasks across five capability domains and runs on 36 production MCP servers in isolated containers. Each task is designed with a structural trigger that makes one of four subtypes the dominant failure risk: Branch Collapse, Unreachable Goal, Tool Misuse, or Context Forgetting. This design enables subtype-level analysis rather than aggregate task success alone. We evaluate fourteen frontier LLM agents from eight vendors on 5,012 trajectories. Current MCP agents remain vulnerable to hallucination-inducing tool environments, achieving an average score of 0.642. Failures concentrate on Unreachable Goal tasks, where the average score drops to 0.458 and 22.1\% of runs pass, indicating that agents often continue issuing tool calls or claim successful completion rather than recognizing infeasible goals. Performance across the four subtypes is largely independent: per-model ranks span on average 6.9 of 13 positions, and Branch Collapse is uncorrelated with Unreachable Goal across models ($r = -0.12$). MCPHallu is released at https://anonymous.4open.science/r/mcp_hallu.

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.

Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions. Recent *thinking with images* approaches mitigate this limitation by allowing models to inspect images through operations such as zooming, cropping, or code execution. While effective, their multi-turn interactions with VLMs introduce redundant token computation and slow inference, and separate region inspection from the reasoning step that depends on it. To address these issues, we propose **Lo**ok Before You **Re**ason (LoRe), an implicit visual thinking framework for efficient visual reasoning within a single-turn VLM interaction. Given an image and a question, LoRe predicts the most relevant image region, aligns it with the corresponding image patch features, and strengthens the associated key-value (KV) cache entries, enabling subsequent decoding to directly exploit the selected visual cues without additional visual inputs or explicit tool calls. We further employ reinforcement learning to guide the VLM toward identifying informative image regions and leveraging them during reasoning. To support fine-grained evaluation of multimodal reasoning, we introduce COCO-ObjQA, a large-scale benchmark for object-centric multimodal reasoning. Experiments on three challenging multimodal benchmarks demonstrate that LoRe consistently improves VLM reasoning performance with low additional inference overhead.

Extremal combinatorics seeks rare mathematical structures that satisfy hard constraints while maximizing a global objective. Most existing neural systems for mathematical discovery train a model to generate complete configurations from scratch. We argue that this framing mismatches the dynamics of effective search: high-quality solutions are more effectively found by preserving a feasible partial core and conditionally regenerating the remaining uncertain objects. We formulate extremal structure discovery as object-level active-core completion: the model is trained to conditionally complete masked objects given a feasible partial core, capturing inter-object relations rather than fitting a distribution over complete configurations. We instantiate this with a full-attention Transformer over object-aligned slots, and propose Inpainting PatternBoost, a variant of PatternBoost. In each iteration, elite constructions from a top pool are partially destroyed by random or structured remasking; the model learns to reconstruct the missing objects, and their completions are reinserted into the pool. We evaluate on the no-three-in-line problem and maximum sum-of-radii circle packing in the unit square, under a single shared abstraction. On no-three-in-line, our framework recovers the optimal $2n$-point configuration for $n$ up to $18$, and reaches $2n-1$ points for $n=19, 20, 25$, outperforming prior neural baselines. On circle packing, where FlowBoost combines neural generation with gradient-based local search, our approach matches or exceeds FlowBoost when paired with the same local search; even without local search, it reaches an average of 88.1% of FlowBoost's reported score across instances.

Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where a single model should answer directly for simple inputs and reason step by step only when needed. The key observation is that offline adaptive thinking is not merely a length-control problem: fast direct answers and slow reasoning traces are typically generated by distinct behavior policies. This dual-reference structure makes standard offline reward optimization, such as Decoupled Generation and Optimization (DGO), mismatched because it assumes a single reference policy and can misweight one mode. We propose \textsc{Dual-Reference DGO}, which models fast and slow responses under a fast/slow mixture reference and derives a reward-weighted offline objective for adaptive VLM reasoning. Our formulation shows that router-free fast/slow switching emerges from competition between fast and slow partition functions, and yields an efficient mode-wise correction for behavior-reference mismatch. Experiments on multiple datasets show that \textsc{Dual-Reference DGO} improves the accuracy--length trade-off over baselines. Ablations show that single-reference DGO tends to collapse to almost always-fast or always-slow behavior, while our dual-reference formulation preserves both modes and learns difficulty-dependent routing.

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.

Unified models (UMs) hold promise for their ability to understand and generate content across heterogeneous modalities. Compared to merely generating visual content, the use of UMs for interleaved cross-modal reasoning is more promising and valuable, e.g., for solving understanding problems that require dense visual thinking, improving visual generation through self-reflection, or modeling visual dynamics of the physical world guided by stepwise action interventions. However, existing UMs necessitate pixel decoding as a bridge due to their disjoint visual representations for understanding and generation, which is both ineffective and computationally cumbersome. In this paper, we introduce LatentUM, a novel unified model that represents all modalities within a shared semantic latent space, eliminating pixel-space mediation for model-internal reasoning over generated visual states. By making generated visual tokens directly interpretable to the model itself, LatentUM enables flexible interleaved cross-modal reasoning and generation. Empirically, LatentUM achieves state-of-the-art performance on the Visual Spatial Planning benchmark, pushes the limits of visual generation through self-reflection and demonstrates world modeling capability by predicting future visual states in the shared semantic latent space.

Humans effortlessly infer 3D structure, such as depth, occlusion, and spatial arrangement, from 2D images and reason about it fluidly. Multimodal large language models still struggle with such spatial reasoning. Current approaches attempt to bridge this gap by injecting explicit 3D priors at inference time or aligning features with 3D-aware teachers during training. While these strategies improve geometric perception, they typically treat 3D knowledge as static representations, rather than enabling the model to reason with 3D cues during inference. We argue that closing this gap requires internalizing 3D knowledge in a form that can participate in the model's reasoning dynamics. To this end, we introduce \textbf{Latent Spatial Reasoning (LSR)}, a framework that distills geometric knowledge into \emph{latent spatial tokens}: geometry‑aware representations that the model actively queries and conditions on to guide its spatial reasoning process. Through hierarchical distillation, LSR transfers 3D knowledge from foundation models, aligns it with the vision-language space, and trains the model to interleave spatial and linguistic reasoning, all without external 3D modules at inference time. Experiments on diverse spatial reasoning benchmarks demonstrate state-of-the-art performance.

Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast model weights before clearing its key-value cache. During the sleep phase, the model performs $N$ offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. This shifts extra computation to the sleep phase while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a more realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration $N$ for our models improves performance, with the largest gains on examples that require deeper reasoning.

Confidence-weighted routing, selective abstention, and ensemble weighting all assume that a model's stated confidence is informative about its capability on the question being asked. They presume functional metacognition, the capacity to assess one's own capabilities, without exercising them, relative to other frontier language models. Aggregate calibration is well studied, with mixed results, but the underlying structure of elicited confidence is less well understood. We decompose binary confidence judgements from 20 frontier Large Language Models (LLMs) across six benchmarks using tetrachoric factor analysis paired with pairwise calibration, asking whether two models that differ in confidence also differ in performance. On factual recall and information retrieval benchmarks the cross-model confidence matrix is approximately rank-one and a single dominant factor captures most of the latent variance. Models retrieving facts share an item-level difficulty axis and differ mainly in their decision thresholds along it. Across all benchmarks the relationship between confidence and performance collapses once items that all models agree on are removed. Inter-model pairwise calibration is small even where statistically significant, and what remains shrinks to nothing once base-rate differences along the shared factor are controlled for. Mathematical reasoning is the apparent exception, but this turns out to be a confound where reasoning models answer questions about their confidence by trying to solve them in their chain of thought, bypassing the sub-symbolic self-knowledge we seek to measure. We find no evidence for significant verbalised individuated metacognition in any tested domain. A trivial logistic regression on surface features of the confidence-judgement reasoning text beats the median confidence model's repeatedly sampled judgements on half the benchmarks, evidence that extracting uncertainty from the reasoning trace recovers signal the binary judgement does not.

Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and latency, yet no existing benchmark systematically studies when a tool call is actually needed. We propose When2Tool, a benchmark of 18 environments (15 single-hop, 3 multi-hop) spanning three categories of tool necessity — computational scale, knowledge boundaries, and execution reliability — each with controlled difficulty levels that create a clear decision boundary between tool-necessary and tool-unnecessary tasks. We evaluate two families of training-free baselines: Prompt-only (varying the prompt to discourage unnecessary calls) and Reason-then-Act (requiring the model to reason about tool necessity before acting). Both provide limited control: Prompt-only suppresses necessary calls alongside unnecessary ones, and Reason-then-Act still incurs a disproportionate accuracy cost on hard tasks. To understand why these baselines fail, we probe the models' hidden states and find that tool necessity is linearly decodable from the pre-generation representation with AUROC 0.89–0.96 across six models, substantially exceeding the model's own verbalized reasoning. This reveals that models already know when tools are needed, but fail to act on this knowledge during generation. Building on this finding, we propose Probe&Prefill, which uses a lightweight linear probe to read the hidden-state signal and prefills the model's response with a steering sentence. Across all models tested, Probe&Prefill reduces tool calls by 48% with only 1.7% accuracy loss, while the best baseline at comparable accuracy only reduces 6% of tool calls, or achieves a similar tool call reduction but incurs a 5× higher accuracy loss. On the real-world Search-o1 agentic benchmark, Probe&Prefill reduces API calls by 20–56% without accuracy degradation.