NeurIPS 2026 Papers

The reasoning capabilities of large language models are reported to drop sharply when they are forced to think in heavily regulated target languages such as Chinese — a phenomenon the literature attributes to data imbalance, multilingual transfer failures, or, more recently, RL-induced cross-lingual collapse. We propose and test an alternative mechanistic account: that the gap reflects RL-induced selection between conditional policies sharing the same parameters, rather than a loss of underlying capability. We construct a model organism for this account. Two policies — identified solely by the language of an unsupervised chain of thought — are installed via SFT on disjoint splits of ARC-Challenge, with one policy trained to answer and the other to refuse on each split. Reinforcement learning with a final-answer reward, restricted to one split, drives the language of reasoning to collapse onto the rewarded policy across both splits, producing a 67% → 10% accuracy collapse on the held-out split despite no reward signal ever observing it. General capability is unaffected: MMLU accuracy is 71.2% before and after RL. A depth-wise activation steering intervention then recovers held-out accuracy from 0% to 85% on the worst-case slice (Chinese-CoT completions) while preserving Chinese as the surface chain-of-thought language and leaving the rewarded policy intact (83% vs. 94% unsteered). The capability was never lost; it was gated. To the extent deployed regulated-language models contain analogous structure, surface evaluations of such models will systematically understate capability hidden behind learned policy gates.

Vision-Language-Action (VLA) models have emerged as a promising paradigm for end-to-end autonomous driving. However, existing reasoning mechanisms still struggle to provide planning-oriented intermediate representations: textual Chain-of-Thought (CoT) fails to preserve continuous spatiotemporal structure, while latent world reasoning remains difficult to use as a direct condition for action generation. In this paper, we propose CoWorld-VLA, a multi-expert world reasoning framework for autonomous driving, where world representations serve as explicit conditions to guide action planning. CoWorld-VLA extracts complementary world information through multi-source supervision and encodes it into expert tokens within the VLA, thereby providing planner-accessible conditioning signals. Specifically, we construct four types of tokens: semantic interaction, geometric structure, dynamic evolution, and ego trajectory tokens, which respectively model interaction intent, spatial structure, future temporal dynamics, and behavioral goals. During action generation, CoWorld-VLA employs a diffusion-based hierarchical multi-expert fusion planner, which is coupled with scene context throughout the joint denoising process to generate continuous ego trajectories. Experiments show that CoWorld-VLA achieves competitive results in both future scene generation and planning on the NAVSIM v1 benchmark, demonstrating strong performance in collision avoidance and trajectory accuracy. Ablation studies further validate the complementarity of expert tokens and their effectiveness as planning conditions for action generation.

Chain-of-thought (CoT) prompting and related strategies that elicit intermediate reasoning traces have fundamentally changed how language models are used, evaluated, and developed. CoT now touches many areas of contemporary research beyond prompting, from advanced test-time inference and model tuning to interpretability, yet much of this work treats CoT traces in task-specific ways, without a shared formal account of what they are or how they compose and support inference. We argue that probabilistic programs — ordinary programs extended with stochastic choices and probabilistic conditioning — provide a natural framework for modeling CoT reasoning. To demonstrate this, we derive a small discrete probabilistic programming language for CoT in which reasoning steps are stochastic choices, branches encode dependencies, trace likelihoods score executions, and answers are return values. Even with a minimal reachability-based semantics, this language enables a first-principles analysis of trace likelihood, exposing why likelihood alone can be a misleading model of reasoning. Motivated by this limitation, we add differentiable operators for composing scores and feedback over CoT traces. The resulting framework links several uses of CoT — scoring, evidence aggregation, feedback propagation, model tuning, and step-level sensitivity analysis — within a single probabilistic program. Through case studies, we show how this approach yields new analysis techniques and test-time inference strategies, offering a promising direction for programming-language-based approaches to language-model reasoning.

Monitoring the chain-of-thought (CoT) of reasoning models is a promising approach for detecting covert misbehavior (i.e., hidden objectives) in code generation tasks. While large models (GPT-5, Gemini-3-Flash) can serve as effective CoT monitors, they are expensive to deploy due to the lengthy reasoning traces and high API cost, emphasizing the need for smaller, cheaper alternatives. Nevertheless, we find that current small models (4B--8B) struggle to detect hidden objectives despite access to the CoT, frequently misattributing them as part of the user query. To address this, we propose a post-training pipeline combining supervised fine-tuning (SFT) and reinforcement learning (RL), where SFT narrows the gap for in-domain tasks by distilling detection behavior from stronger monitors, and RL on hard and subtly crafted hidden objectives helps the model generalize to out-of-domain monitoring tasks. To validate this generalization, we evaluate under a realistic threat model motivated by practical supply-chain attacks, where the adversary is a third-party LLM router injecting hidden objectives into code-generation requests through either prompt manipulation or code manipulation attacks. To push beyond objectives that large monitors already saturate, we also introduce four new challenging tasks even for strong monitors. Finally, we introduce CoT-Guard, a 4B-parameter monitor that demonstrates superior generalization performance under both prompt and code manipulation attacks, achieving a $\text{Gmean}^2$ (i.e., TNR$\times$TPR) of 75\% and outperforming GPT-5.4 (56\%), GPT-5-mini (41\%), and Qwen3-32B (54\%), while closing the gap to Gemini-3-Flash (83\%). These results demonstrate that \modelname provides a practical and cost-effective user-side defense, substantially improving hidden-objective detection while avoiding the deployment cost of large monitors.

Large language models have made knowledge-graph question answering more flexible, yet complex multi-hop questions often require evidence that becomes identifiable only during reasoning. Starting from the question alone, narrow retrieval can miss bridge facts, while broad expansion injects distractors and worsens long-context interference. The key bottleneck is dynamic evidence identification: the system must revise a compact evidence view as reasoning exposes unsupported facts. We formulate this process as partially observable control over evidence and propose **CoRE**, a residual-coupled framework in which reasoning trajectories and evidence subgraphs evolve together. CoRE decomposes each trace into answer-critical claims, treats unsupported claims as residual signals, and uses them to guide targeted retrieval and budgeted subgraph projection. The same formulation exposes a learnable projection decision: after residuals retrieve candidate facts, **CoRE-RL** learns from rollout feedback which facts should remain visible for the next reasoning pass. Across WebQuestionsSP and Complex WebQuestions, the CoRE family improves answer accuracy, evidence sufficiency, and budget robustness, with CoRE-RL providing further gains through more effective retained-subgraph selection.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose CoQuant, an Ultra-low-bit KV Cache quantization method that estimates *attention-aware covariance* structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable CoQuant system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, CoQuant reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale CoQuant to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, CoQuant remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, CoQuant reduces KV-cache memory by approximately 8×, improves throughput by up to 7× at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3× over BF16 due to reduced memory bandwidth overhead.

Large reasoning models (LRMs) improve task performance by generating longer intermediate reasoning traces, but this substantially increases inference costs and makes low-precision deployment increasingly important. However, aggressive quantization often performs poorly on reasoning tasks. Beyond error accumulation across autoregressive generation, we observe that activations exhibit significantly higher variability during the early decoding steps of thinking, making conventional calibration and uniform quantization less effective. To address these reasoning-specific quantization challenges, we propose \textbf{ClusQuant}, a clustering-based LUT quantization framework for low-precision large reasoning models. ClusQuant strengthens outlier mitigation through rotation and scaling transformations, uses clustering-based representations to better fit non-uniform activation distributions, and replaces fixed codebook sizes with an adaptive clustering strategy that allocates codebook capacity according to layer-wise quantization difficulty. Guided by our observation of reasoning-time activation behavior, ClusQuant further improves calibration sample selection to better match actual decoding statistics. Combined with our customized LUT CUDA kernel, ClusQuant delivers substantial accuracy improvements on reasoning benchmarks while achieving $2.85\times$ speedup.

Large language models (LLMs) have demonstrated impressive reasoning capabilities, but scaling their performance often relies on massive reasoning datasets that are computationally expensive to train on. Existing data selection methods aim to curate smaller, high-quality subsets but often rely on costly external models or opaque heuristics. In this work, we shift the focus from external heuristics to the model's internal mechanisms. We find that complex reasoning tasks consistently activate a sparse, specialized subset of attention heads, forming core reasoning circuits. Building on this insight, we propose CircuitSeer, a novel data selection method that quantifies the reasoning complexity of data by measuring its influence on these crucial circuits. Extensive experiments on 4 models and 9 datasets demonstrate CircuitSeer's superiority. Notably, fine-tuning Qwen2.5-Math-7B on just 10\% of data selected by our method achieves a 1.2-point gain in average Pass@1 over training on the full dataset, highlighting its efficiency and effectiveness.

Cinematic understanding requires more than recognizing visible events: models must acquire dialogue evidence, bind it to speakers, and use it to infer narrative meaning. Existing video benchmarks evaluate pieces of this problem, including video question answering, audio-visual reasoning, and temporal grounding, but rarely test whether these abilities form a coherent cinematic evidence chain within the same clips. As a result, a model may answer a movie question correctly while failing to localize the relevant utterance, identify the speaker, or integrate speech with visual context. We introduce **CineMME**, a human-verified benchmark designed to diagnose speech-grounded cinematic reasoning, supported by approximately $1,600$ expert person-hours of annotation. CineMME contains $975$ curated movie clips and two linked tracks: *CineMME-Grounding*, with $13,554$ timestamped speech segments and $33,083$ actor face boxes for temporal dialogue localization, transcription, and active-speaker grounding; and *CineMME-Reasoning*, with $1,600$ multiple-choice and $316$ open-ended questions spanning perception, audio, narrative, social, and cinematic understanding. We evaluate $20$ MLLMs on reasoning and $10$ omni-modal models on grounding. The results reveal a pronounced who-what-when decoupling: Gemini-3 Pro localizes dialogue in time 68.33% tIoU) and partially transcribes it (WER 29.00%), yet reaches only 13.24% sIoU for active-speaker grounding while achieving 65.88% reasoning accuracy. Modality ablations further show that vision is not uniformly beneficial: it improves perception-heavy questions but can hurt dialogue-centric reasoning. CineMME provides a compact diagnostic testbed for measuring whether multimodal video models truly connect dialogue, speakers, visual evidence, and narrative meaning, rather than succeeding through disconnected cues.

Charts convey dense quantitative and relational information, yet general chart parsing remains challenging due to fragmented evaluation and limited structural reasoning. Existing benchmarks focus on narrow sub-tasks with inconsistent output formats (e.g., Markdown, CSV, JSON, SVG, code), hindering fair comparison and often overlooking real-world scenarios such as printed or hand-drawn photos. Meanwhile, end-to-end multimodal models tend to learn shallow pixel-to-text mappings without explicit structure understanding, making them fragile under visual variations. To address these issues, we present a unified framework that connects evaluation and modeling. On the evaluation side, we introduce **ChartArena**, a comprehensive bilingual benchmark covering eight chart families, including both numeric charts (bar, line, pie, radar, box plot, combination) and diagrammatic structures (flowchart, mind map). The dataset is constructed via a human-agent collaborative annotation pipeline, where model-assisted drafts are refined through multi-stage human verification to ensure structural consistency and annotation reliability. Furthermore, we design a format-agnostic evaluation protocol that maps different outputs into two canonical semantic spaces: a normalized triple view and a directed graph view, and evaluates them with structure-aware metrics. On the modeling side, we propose **Chart Atomic Primitives (CAP)**, a lightweight semantic scaffold that introduces structural inductive bias during supervised fine-tuning. By guiding the model to infer structure before generating outputs, CAP leads to more consistent and robust parsing behavior. Extensive experiments show that models trained with CAP achieve reliable performance on both rendered charts and real-world images. Code, benchmark, and models will be made publicly available.

Spatial intelligence requires foundation models to maintain coherent spatial state across interactions with the physical world. However, existing data-centric approaches typically treat spatial reasoning as independent question-answer instances, enabling shortcut-based answering and providing limited supervision for persistent spatial understanding. To address this, we introduce ChainSpace, a chained-reasoning paradigm that structures spatial reasoning as a state-preserving multi-round process. In this paradigm, spatial questions are organized into logically constrained and jointly consistent chains, where later questions depend on spatial constraints established in earlier rounds. Following this principle, we instantiate ChainSpace-Bench, a manually annotated real-world multi-round benchmark with a Chain-Aware Metric, and ChainSpace-Pipeline, a simulator-based chain-structured supervision generation framework for spatial intelligence training. Experiments show that ChainSpace-Bench exposes chain-level failures that are not captured by isolated question accuracy. Additionally, with a relatively small amount of simulator-generated chained data, models trained by ChainSpace-Pipeline achieve the best performance among open-source models on ChainSpace-Bench and transfer competitively to multiple external spatial intelligence benchmarks. These results establish ChainSpace as an effective paradigm for more faithful evaluation and more data-efficient learning of spatial intelligence.

*Chains‑of‑thought* (CoT) allow language models to verbalise multi‑step rationales before producing their final answer. While this technique often boosts task performance and offers an impression of transparency into the model's reasoning, this position paper argues that rationales generated by current CoT techniques can be misleading and are *neither necessary nor sufficient* for trustworthy interpretability. We analyze *faithfulness* as whether CoTs are not only human-interpretable, but also reflect the model’s underlying reasoning well enough to support responsible use. By synthesizing evidence from prior work, we show that verbalised chains are frequently *unfaithful*, diverging from the true hidden computations that drive a model's predictions. As CoT is increasingly relied upon in application domains, including high-stakes ones such as medicine, law, and autonomous systems, we further argue that this unfaithfulness is not taken seriously enough in these settings: our analysis of 1,000 recent CoT-centric papers finds that approximately 25\% explicitly treat CoT as an interpretability technique, including papers in high-stakes domains which heavily hinge on such interpretability claims. Building on prior work, we make three proposals: (i) avoid treating CoT as being sufficient for interpretability without additional verification, while continuing to use CoT for its communicative benefits, (ii) adopt rigorous methods that assess faithfulness for downstream decision-making, and (iii) develop causal validation methods to ground explanations in model internals.

Cellular networks are critical infrastructure supporting billions of worldwide users and safety- and mission-critical services. Vulnerabilities in cellular networks can therefore cause service disruption, privacy breaches, and broad societal harm, motivating growing efforts to analyze 3GPP specifications that define required device and operator behavior. However, existing benchmarks for this domain share three systematic deficiencies: they rely on LLM-generated entries without expert verification, they do not require faithful interpretation of normative language, reasoning across cross-referenced clauses, and verifiable conclusions grounded in multimodal evidence such as tables and figures, and they do not evaluate security reasoning over end-to-end protocol behavior. To address these gaps, we present CellularS2-Bench, a staged benchmark. CellularS2-Bench covers core control-plane specifications, processing over 1,760 paragraphs, 13,800 sentences, 424 figures, and 340 tables. It comprises 7,851 questions and instances derived from 3GPP specifications and 43 real-world vulnerabilities across three progressive stages. Beyond newly constructed tasks, CellularS2-Bench integrates and corrects subsets from 7 prior datasets to broaden task coverage. To establish a reproducible baseline on CellularS2-Bench, we introduce CellularS2-ARI, a unified Adapt-Retrieve-Integrate framework for systematic understanding and specification-driven security analysis of 3GPP specifications. Together, they establish an accessible and reproducible foundation for quantifying progress in specification understanding and security reasoning in cellular network security.

Humans can look at a static scene and instantly predict what happens next --- *will moving this object cause a collision?* We call this ability Causal Spatial Reasoning. However, current multimodal large language models (MLLMs) cannot do this, as they remain largely restricted to static spatial perception, struggling to answer ``what-if'' questions in a 3D scene. We introduce **CausalSpatial**, a diagnostic benchmark evaluating whether models can anticipate consequences of object motions across four tasks: Collision, Compatibility, Occlusion, and Trajectory. Results expose a severe gap: humans score 84\% while GPT-5 achieves only 54\%. Why do MLLMs fail? Our analysis uncovers a fundamental deficiency: models over-rely on textual chain-of-thought reasoning that drifts from visual evidence, producing fluent but spatially ungrounded hallucinations. To address this, we propose the Causal Object World Model (COW), a framework that externalizes the simulation process by generating videos of hypothetical dynamics. With explicit visual cues of causality, COW enables models to ground their reasoning in physical reality rather than linguistic priors. Code and dataset are publicly available.

Recently, video generation models have achieved remarkable progress in global visual fidelity. However, when synthesizing complex dynamic content, they frequently produce **sparse subtle anomalies**. Existing video reward models primarily emphasize overall visual quality and text alignment; yet as generation quality improves, these coarse-grained defects become less frequent, while the sparse structural, temporal, and physical anomalies become an increasingly important bottleneck. Therefore, in this paper, we propose Concentrate and Concentrate (CaC), a coarse-to-fine anomaly reward model based on Vision-Language Models. During inference, it first conducts a global temporal scan to anchor anomalous time windows, then performs fine-grained spatial grounding within the localized interval, and finally derives robust judgments via structured spatiotemporal Chain-of-Thought reasoning. To equip the model with these capabilities, we construct the first large-scale generated video anomaly dataset with per-frame bounding-box annotations, temporal anomaly windows, and fine-grained attribution labels. Building on this dataset, we design a three-stage progressive training paradigm. The model initially learns spatial and temporal anchoring through single- and multi-frame supervised fine-tuning, and then is optimized by a reinforcement learning strategy based on two-turn Group Relative Policy Optimization (GRPO). Beyond conventional accuracy rewards, we introduce Temporal and Spatial IoU rewards to supervise the intermediate localization process, effectively guiding the model toward more grounded and interpretable spatiotemporal reasoning. Extensive experiments demonstrate that CaC can stably ``concentrate'' on subtle anomalies, achieving a 25.7\% accuracy improvement on fine-grained anomaly benchmarks and, when used as a reward signal, CaC reduces generated-video anomalies by 11.7\% while improving overall video quality.

Code agents must both reason over long-horizon repository state and obey strict tool-use protocols. In paired Instruct/Thinking checkpoints, these capabilities are complementary but misaligned. The Instruct model is concise and tool-disciplined, whereas the Thinking model offers stronger planning and recovery behavior but often over-deliberates and degrades agent performance. We present CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing), a training-free parameter-editing method that treats the Thinking–Instruct delta as a directional pool of candidate reasoning edits for the Instruct backbone. CRANE combines magnitude thresholding to denoise the delta, a Conservative Taylor Gate to retain edits that are jointly beneficial for reasoning transfer and tool-use preservation, and Graduated Sigmoidal Projection to suppress format-critical update directions. By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass@1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass@1/pass@5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.

Large Language Models (LLMs) are presently well assessed by expert–level mathematics and physics benchmarks which focus on theoretical, symbolic reasoning. However, we lack experimental benchmarks that evaluate data analysis in frontier research problems, a capability that future domain–specific foundation models must have. To this end, we present CME–SpectrumBench, an expert–designed benchmark consisting of 190 questions that measures how reliably frontier LLMs analyze research–level spectral scientific data in condensed matter experiments (CME). Using simulated CME spectral data, we prompt models to identify structures from 2D intensity arrays in the face of noise and limited instrumental resolution, a test of dense context reasoning. Models are scored by how accurately their numerical answers approach the ground truth. Under direct query, scores generally pale in comparison to handwritten code due to clear failure modes: models can find simple features but struggle to trace patterns or compute derived quantities, and performance degrades with increased data resolution and noise/convolution levels. LLM performance becomes comparable to handwritten code by allowing tool use (access to a Python interpreter) with/without ReAct prompting, in effect mapping questions onto agentic code generation problems. Our findings show that frontier models can serve as research assistants by interfacing with data through code, but still face difficulties when processing information directly from large datasets — a key barrier future foundation models must overcome.

Structured road understanding of lane geometry, topology, and traffic element relationships is foundational to safe autonomous driving. While vision-language models (VLMs) offer promising semantic flexibility, they lack the geometric and relational grounding required for precise road reasoning. Conversely, traditional modular systems, e.g., HD maps and topological road graphs, provide structural precision but remain semantically rigid. To bridge this gap, we introduce the Combined Road Substrate (CRS), a graph-grounded framework that makes geometric road structure and open-vocabulary semantics jointly executable in a single representation. CRS enables the automatic generation of compositionally complex and linguistically varied question-answer pairs via recursive graph queries, augmented with a ``grounding for free'' mechanism that ensures logical traceability to specific map elements, and procedurally extracted chain-of-thought supervision traces. We demonstrate that state-of-the-art VLMs - including large, closed-source models - struggle significantly with structured road reasoning, yet training a small 2- or 4-billion-parameter model with as few as 20 to 80 CRS-enriched scenes yields stable gains in compositional reasoning tasks of varying depth. Analysis of model behavior via verifiable reasoning traces reveals a systematic shift in failure modes: whereas baseline models fail at relational scene understanding, CRS-trained models reduce failures to attribute recognition, suggesting that the primary bottleneck in road understanding is not model scale, but the absence of structured supervision.

Classical compute-optimal scaling laws assume an unbounded supply of fresh pretraining data, yet pretraining is increasingly entering the regime where compute is growing faster than high-quality data. We propose *Compute-Data (CD) scaling laws*, a unified framework that bridges *compute-optimal* scaling, in which data scales freely with compute, and *data-optimal* scaling, in which the corpus is fixed and compute can grow unbounded. CD scaling extends classic scaling by introducing a *token effectiveness function* that quantifies how much a *derived* token, produced for instance by multi-epoch repetition or paraphrasing, is worth relative to a fresh one, ranging from a perfect substitute to no value at all. Fitting $\eta$ for two data expansion strategies (multi-epoch repetition, paraphrasing) across model sizes ranging from 14M to 600M parameters on the Dolma-3 corpus, we find that it is far from constant: it depends jointly on model size, the tokens-per-parameter ratio, and the amount of derived data, and saturates as the corpus is expanded. The functional form of $\eta$ implies that substituting compute for data *diminishes* with both model size and data availability, and partitions training into three operational regimes: compute-bound, data-bound, and model-bound, showing that classic compute-optimal allocation is suboptimal across most of the practically relevant regime.