NeurIPS 2026 Papers

Achieving peak GPU kernel performance increasingly relies on architecture-specific optimizations targeting new hardware features. While AI coding agents show promise in generating performant kernels, they lack the necessary context to effectively implement and stack hardware-specific optimizations, especially on newer GPU architectures. We propose Hawkeye (Hardware-Aware Kernel Optimization), an open-source framework that grounds autonomous kernel generation in a minimal and comprehensive taxonomy with only one unit test per optimization strategy per target architecture. Supporting a new accelerator therefore requires only 10 expert-written unit tests per architecture (one per recurring optimization strategy) that generalize across downstream workloads, rather than hand writing a new kernel for each workload and precision. Hawkeye effectively scales the test-time compute of coding agents with this minimal expert supervision to enable kernel generation that consistently leverages hardware-specific features, approaching and even surpassing expert-written PyTorch or Triton in BF16 and emerging low precision (FP8, NVFP4, MXFP4) across Ampere, Hopper, Blackwell, and MI350 GPUs. Hawkeye demonstrates that minimally supervised coding agents can exploit architecture-specific hardware features and reduce the overhead of supporting emerging hardware accelerators.

Long-video understanding in VLMs is bottlenecked by a single monolithic forward pass over thousands of frames at quadratic attention cost. A common mitigation is to first $select$ a small subset of informative frames before the forward pass; common for training-free selectors via auxiliary encoder-space similarities. Such signals are capped by contrastive pretraining, which usually fails on reasoning-heavy queries (negation, cross-frame counting, holistic summarization). We propose $\textbf{GridProbe}$, an efficient training-free posterior-probing inference paradigm that scores evidence in $\textbf{answer space}$ using a frozen VLM's own reasoning and then selects question-relevant frames $\textbf{adaptively}$, resulting in sub-quadratic attention cost with little to no accuracy loss. We arrange frames on a $K{\times}K$ grid and run lightweight row $R$ and column $C$ probes, where each probe reads its peak posterior as a query-conditioned confidence. The outer product of $R$ and $C$ yields an interpretable $\textbf{importance map}$ whose skewness and kurtosis drive $\textbf{Shape-Adaptive Selection}$, a closed-form rule that reliably replaces the fixed frame budget $M$ with a per-question $M_{\mathrm{eff}}$. We show empirically that $M_{\mathrm{eff}}$, surprisingly, tracks intrinsic question difficulty without ever seeing the answer, a sign of test-time adaptive compute. On Video-MME-v2, $\textbf{GridProbe}$ matches the monolithic baseline within $1.6$ pp Avg Acc at $3.36\times$ TFLOPs reduction, while on LongVideoBench it Pareto-dominates the baseline ($+0.9$ pp at $0.35\times$ compute). Because the selector and QA models can be decoupled, pairing a small 2B selector with a stronger 4B or 8B QA is strictly Pareto-dominant over the 2B monolithic baseline (up to $+4.0$ pp at $0.52\times$ compute, on average), with no retraining. Finally, the interpretability of the importance maps opens future avenues for behavioral diagnostics, grounding, and frame-selection distillation.

Memory consumption of the key-value (KV) cache remains a central bottleneck for long-context LLM inference. We introduce PRKV, a training-free KV eviction method that estimates multi-hop global token importance by applying Personalized PageRank over attention-induced transition graphs. This enables importance propagation beyond recency or instantaneous attention, without modifying the model. We benchmark PRKV on long-context understanding, sparse retrieval, and chain-of-thought reasoning tasks using multiple LLM families. Under moderate to aggressive compression ratios, including up to 90\%, PRKV achieves competitive task performance against established eviction baselines while reducing peak memory usage with low end-to-end overhead. These results demonstrate that multi-hop global influence estimation yields a favorable accuracy--memory trade-off for scalable long-context inference.

Chain-of-thought (CoT) reasoning enables large language models (LLMs) to solve complex problems by generating intermediate reasoning steps. While much attention has been paid to the length and content of these reasoning chains, far less is known about their internal geometry. We study the *geometry* of CoT trajectories in the hidden state space of transformer models, formalizing each reasoning chain as a discrete curve in $\mathbb{R}^d$ and characterizing it through spectral, positional, and kinematic geometric functionals. We introduce the effective dimension $d_\rho$ as a measure of trajectory complexity and show theoretically that trajectories with flatter eigenvalue spectra correspond to harder tasks, as they explore more of the hidden dimensions. Lastly, we explore how kinematic features of the trajectory, mean position, positional dispersion, initial hidden states, mean velocity, mean speed, and speed dispersion, can be used to predict solution correctness before generation is complete, and may inform future early-stopping strategies. Experimentally, on mathematical reasoning problems from the MATH500 dataset, $d_\rho$ achieves 0.93 AUC in distinguishing easy from hard problems, while kinematic features show promise for predicting correctness from only the first 20% of generated tokens. These correctness signatures transfer across questions of varying difficulty, establishing that the *shape* of a model's internal reasoning trajectory is a principled window into both task hardness and solution quality.

Test-time scaling (TTS) improves large language model (LLM) reasoning by spending additional inference-time compute, most commonly through width-heavy repeated sampling and aggregation. However, this strategy scales cost almost linearly with the number of calls and can remain brittle when the sampled trajectories share the same failure mode. We introduce Maze, a test-time scaling framework that shifts computation away from repeated target-model calls and toward LLM-free planning over ordered few-shot exemplars. The central idea is to treat an ordered exemplar sequence as a prompt trajectory embedded in a latent geometric space. Instead of asking the LLM for many independent attempts, Maze first ranks candidate prompt trajectories using cached encoder representations and a path scorer, ${\it i.e.}$, a lightweight learnable metric, then spends one call (or a few calls) only on the most promising trajectories. This design preserves black-box deployability, since the planner is decoupled from the target LLM and can be implemented with a separate frozen encoder. Across various challenging reasoning benchmarks, single-call Maze consistently outperforms strong single-call prompt-optimization baselines, and MultiMaze recovers a large fraction of width-heavy TTS gains with substantially fewer computational cost measured by number of LLM calls and latency. We further present a theoretical perspective showing when prompt-path ranking can recover much of the benefit of best-of-$N$ style scaling: if high-utility prompt trajectories are sufficiently rankable and sufficiently covered by the explored pool, a deeper planner-side search can substitute for a substantial amount of width-heavy sampling.

Large language models solve complex problems by generating lengthy chains of explicit reasoning tokens. While effective, this makes reasoning expensive, length-sensitive, and constrained to (discrete) natural language. While latent reasoning offers a continuous alternative, determining useful structures for intermediate latent states is an open challenge. In this paper, we formulate latent reasoning as a geometric path-approximation problem within the model’s pretrained token-embedding space. We introduce Geometric Latent Reasoning (GLR), which uses a lightweight transition head to predict iterative direction updates in embedding space. Using textual chain-of-thought traces as anchors, GLR learns to approximate discrete reasoning trajectories while permitting continuous deviations from exact token embeddings. Evaluations on mathematical reasoning benchmarks using Qwen3 models reveal an emergent phenomenon: geometric latent reasoning induces substantially shorter generations without an explicit length objective. By replacing early explicit reasoning with continuous latent steps, models often reach correct answers using substantially fewer total generation steps. These findings suggest that continuous trajectories act as compact intermediate reasoning states, exposing a new tradeoff between latent computation budget, output length, and accuracy.

How do transformer language models memorize factual associations? A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, \emph{geometric} form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices: subject embeddings encode \emph{linear superpositions} of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute via ReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting---chains of relational queries such as ``Who is the mother of the wife of $x$?''---providing constructions with and without chain-of-thought that exhibit a provable capacity--depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with precisely the predicted structure. Once trained, the MLP transfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.

Large Multimodal Models (LMMs) often struggle with geometric reasoning due to visual hallucinations and a lack of mathematically precise Chain-of-Thought (CoT) data. To address this, we propose the GeoSym Engine, an automated and scalable neuro-symbolic framework. By leveraging a type-conditional grammar and an analytic SymGT Solver, it derives exact symbolic ground truths and seamlessly integrates with a robust rendering pipeline to produce high-precision geometric diagrams. Using this engine, we construct GeoSym127K, a difficulty-stratified dataset featuring 51K high-resolution images, 127K questions with symbolic ground truths, and 55K answer-verified CoT QA pairs. We also introduce GeoSym-Bench, an expert-curated suite of 511 complex samples for rigorous evaluation. Through extensive supervised fine-tuning (SFT), we demonstrate that GeoSym drives concentrated improvements specifically on diagram-dependent and multi-step geometry tasks. Our Qwen3-VL-8B model gains an absolute +22.21% on the MathVerse Vision-Only subset and reaches 61.52% (+6.19% improvement) on WeMath, mitigating long-horizon logic fragmentation and outperforming advanced closed-source models like Doubao-1.8. Furthermore, applying Reinforcement Learning with Verifiable Rewards (RLVR) via GRPO reveals that initializing from structural SFT checkpoints substantially elevates the performance ceiling over zero-shot RL. Driven by deterministic exact-match signals, this showcases the robust scaling potential of our verifiable reasoning synthesis. Datasets and code are available at https://huggingface.co/datasets/Tomie0506/GeoSym127K and https://github.com/Tomie56/GeoSym127K

Test-Time Scaling (TTS) improves large language models (LLMs) by allocating additional computation during inference, typically through parallel, sequential, or hybrid scaling. However, prior studies often assume fixed collaboration architectures (e.g., topologies) and single-model usage, overlooking that optimal architectures and model combinations can vary across tasks. Therefore, we study the novel problem of \textit{searching for compute-optimal model combinations and architectures in TTS under a fixed budget.} We formalize it as a multi-LLM collaboration graph, where nodes encode roles, and LLM model assignments, and edges capture information flow. This problem is challenging because (i) the combinatorial search space is prohibitively large, and (ii) task-specific requirements demand tailored designs. To this end, we formulate the problem as probabilistic graph optimization and, via pilot experiments, derive three empirical insights into TTS collaboration graphs. Guided by these insights, we propose Agent-REINFORCE, an LLM-agent-augmented framework that mirrors the REINFORCE pipeline by mapping \emph{sampling–gradient–update} to \emph{sampling–feedback–update}, where feedback serves as a textual gradient to update the probabilistic graph and efficiently search for optimal multi-LLM collaboration graphs. Experiments show that Agent-REINFORCE outperforms both traditional and LLM-based baselines in sample efficiency and search performance, and effectively identifies optimal graphs under joint objectives of accuracy and inference latency.

Multimodal Chain-of-Thought (CoT) reasoning has significantly advanced large models, yet forcing continuous visual evidence into discrete text or fixed tool calls inevitably sacrifices fine-grained detail. Recent paradigms attempt to reason directly within a continuous latent space; however, supervising latent slots in isolation fails to account for their collective geometry, resulting in a fragmented latent space that lacks structural guidance. To ensure structural integrity across training stages, we propose GeLVR(Geometry-Consistent Latent Visual Reasoning), a framework that establishes geometric consistency as a governing principle: the relational structure is explicitly aligned during SFT and actively preserved throughout RL, anchoring both stages to a common geometric scaffold. To operationalize this paradigm during the reward-driven phase, we further introduce GePO (Geometry-preserving Policy Optimization), a policy-optimization algorithm tailored for the hybrid discrete-continuous action spaces of latent reasoning. GePO employs a spherical von Mises-Fisher (vMF) policy to respect the decoder's inherent geometry and integrates the geometry-preserving regularizer directly into the RL objective. This ensures that reward optimization and structural integrity advance jointly rather than in tension. Extensive experiments across multiple visual reasoning benchmarks demonstrate that GeLVR consistently outperforms state-of-the-art baselines, particularly in high-resolution and fine-grained perception tasks. Comprehensive ablations further validate the necessity of each geometry-consistent component in establishing stable and coherent latent reasoning. Code is available in the supplementary materials.

Open-source native GUI agents still lag behind closed-source systems on long-horizon tasks. One reason is the direct reuse of generic post-training pipelines that ignore GUI-specific failure modes: standard supervised fine-tuning (SFT) with long chain-of-thought (CoT) reasoning often degrades grounding, and stronger offline optimization in RL does not necessarily translate to better online performance. This offline-to-online mismatch arises in part from \emph{partial verifiability}: multiple actions may validly advance a task, but supervision typically marks only one demonstrated action as correct, creating reward ambiguity. We present \textbf{GUI-Libra}, a data-efficient post-training recipe for reliable reasoning-and-acting in native GUI agents. GUI-Libra combines a construction and filtering pipeline for a curated 81K GUI reasoning dataset, \emph{action-aware supervised fine-tuning} that mixes reasoning-then-action and direct-action supervision with action-aware token reweighting, and KL-constrained RL with success-adaptive scaling to improve offline-to-online predictability under ambiguous rewards. Across web and mobile benchmarks, GUI-Libra consistently improves both step-wise accuracy and end-to-end task completion. GUI-Libra-4B and GUI-Libra-8B improve their base models by +15.6\% and +12.2\% on AndroidWorld, +4.0\% and +8.7\% on Online-Mind2Web, and +12.5\% and +11.3\% on WebArena-Lite-v2. These results show that careful reasoning data curation and tailored post-training can substantially improve long-horizon task solving without costly online data collection. We release our dataset, code, and models to support future research.

Oracle-guided biological sequence design must improve predicted function without moving outside the sequence distribution where the oracle is trustworthy. We introduce Generative Population Annealing GPA, a test-time sampler for sequence design with a frozen generator and frozen oracle. GPA instantiates annealed Sequential Monte Carlo at population scale: particles are initialized from a pretrained sequence prior, reweighted by an oracle reward tilt, selectively upsampled as effective sample size falls, mutated with the pretrained generator, and returned as the design pool. The base sampler targets a reward-tilted prior; practical variants shape the proposal or modify the objective to trade activity, specificity, fidelity, and diversity. Across enhancer and promoter benchmarks, GPA scales to thousands of sequences and is competitive with inference-time samplers, gradient-based editors, tree-search diffusion, and RL-fine-tuned generators. The same inference loop is evaluated with masked discrete-diffusion and autoregressive DNA generators. GPA reaches high predicted activity while preserving strong motif fidelity and model-based likelihood, although specialized baselines retain higher diversity or stronger $k$-mer fidelity in some settings. Cross-oracle and sequence-level audits suggest fewer obvious reward-hacking pathologies, including shorter homopolymer runs than CTRL-DNA in the HepG2 audit (5.1 bp mean maximum run versus 20.9 bp), but all validation remains computational.

Large language models (LLMs) exhibit pronounced position bias in long-context retrieval, systematically prioritizing information location over relevance. Existing mitigations modify positional encodings or attention internals, but in the common output-only deployment setting these are inapplicable. We introduce GOLD PANNING, a Bayesian framework for inference-time active search that (i) reorders documents to concentrate high-belief items in highly diagnostic positions (signal anchoring) and (ii) updates relevance beliefs from model outputs. Unlike active learning, which prioritizes uncertainty reduction, GOLD PANNING exploits anchoring---once flagged, keep it in sight---to preserve weak cues. A greedy assignment derived from the model's diagnosticity profile provably identifies a target among $N$ documents with high probability in $O(\log N)$ rounds. Across open-weight and closed-source models on multi-document QA, GOLD PANNING matches Permutation Self-Consistency's $F_1$ with $30$--$65\%$ fewer queries and remains effective under calibration mismatch, indicating that coarse positional ordering alone drives the gains. These results show that inherent model biases need not be failures, but can serve as exploitable signals.

Knowledge distillation from large reasoning models (LRMs) often uses raw teacher trajectories as supervision. However, this practice creates a mismatch between the goal of distillation and the nature of teacher trajectories. Advanced LRMs typically improve their reasoning capabilities through reward-driven post-training, such as RLHF. This training optimizes them as strong solvers, rather than as teachers that produce reasoning processes smaller models can easily learn. Since smaller models have different "cognitive capacities" compared to their larger teachers, directly imitating raw teacher trajectories can sometimes be ineffective and often requires a substantial number of high-quality samples. To bridge this mismatch, we propose **C**ognition-**A**ligned **R**easoning **D**istillation (**CARD**), an RL-based framework that transforms solver trajectories into teaching trajectories aligned with the student’s cognition. Avoiding expensive resampling, our framework directly transforms existing trajectories. We optimize CARD with two reward signals: (1) a Distribution Alignment Reward that matches the student’s predictive distribution, and (2) a Logical Coherence Reward that reduces misleading failed reasoning steps. Comprehensive evaluations across multiple student models and settings demonstrate that CARD outperforms distillation from raw teacher trajectories in in-domain settings while achieving substantial gains in out-of-distribution performance.

Generating humorous memes is a challenging multimodal task that moves beyond direct image-to-caption supervision. It requires a nuanced reasoning over visual content, contextual cues, and subjective humor. To bridge this gap between visual perception and humorous punchline creation, we propose HUMOR, a novel framework that guides VLMs through hierarchical reasoning and aligns them with group-wise human preferences. First, HUMOR employs a hierarchical, multi-path Chain-of-Thought (CoT): the model begins by identifying a template-level intent, then explores diverse reasoning paths under different contexts, and finally anchors onto a high-quality, context-specific path. This CoT supervision, which traces back from ground-truth captions, enhances reasoning diversity. We further analyze that this multi-path exploration with anchoring maintains a high expected humor quality, under the practical condition that high-quality paths retain significant probability mass. Second, to capture subjective humor, we train a pairwise reward model that operates within groups of memes sharing the same template. Following established theory, this approach ensures a consistent and robust proxy for human preference, even with subjective and noisy labels. The reward model then enables a group-wise RL optimization, providing a theoretical guarantee for monotonic improvement within the trust region. Extensive experiments show that HUMOR empowers various VLMs with superior reasoning diversity, more reliable preference alignment, and higher overall meme quality. Beyond memes, our work presents a general paradigm for open-ended, human-aligned multimodal generation, where success is guided by comparative judgment within coherent output groups.

Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, yet the internal mechanisms driving these multi-step processes remain opaque. Existing mechanistic interpretability approaches often rely on superficial text-pattern co-occurrences or capture only short-term effects, struggling to trace the long-horizon, outcome-driven influence of internal components such as neurons or sparse features in multi-step reasoning. In this paper, we introduce Integrated Policy Gradient (IPG), a novel, training-free framework that brings policy-based attribution into mechanistic interpretability for LLM reasoning. Grounded in *outcome-oriented* and *sequential-influence-aware* principles, it backpropagates outcome-level, non-differentiable behavioral signals (e.g., reasoning outcome correctness) through entire inference trajectories, utilizing path integration to isolate outcome-relevant mediators at component level for reliable attribution. This provides a linear approximation of Causal Mediation Analysis (CMA) in reasoning. Empirical evaluations demonstrate that IPG achieves precise localization and reveals mechanistic insights by identifying sparse, concentrated subsets of internal components associated with specific reasoning behaviors. Our method also enables modulation of reasoning capabilities, offering a powerful and reliable framework for understanding LLM reasoning.

Large language models (LLMs) generally lack deep causal reasoning capabilities in temporal reasoning. Inspired by document layout reordering, this paper proposes a paradigm shift from discrete narratives to causal topologies. It constructs two architectures: a training-free neuro-symbolic framework, temporal causal reasoning agent (TCR-Agent), which achieves interpretable causal deduction and counterfactual truncation through external causal graph instantiation and a symbolic intervention engine; and an end-to-end model, temporal causal reasoning former (TCR-Former), which internalizes these mechanisms into the Transformer attention space via a causal-topological attention mask and temporal span biases. Experiments on the TempoCausal benchmark demonstrate that TCR-Agent comprehensively surpasses existing training-free baselines. At the same time, TCR-Former achieves an accuracy 16.5% higher than that of state-of-the-art (SOTA) methods, outperforming large-scale models such as GPT-5.4 and DeepSeek-V3.2 with only approximately 8.6B parameters. Further analysis reveals a structure-reasoning disconnection phenomenon in existing methods: external symbolic constraints and chain-of-thought (CoT) reasoning can improve reasoning steps, yet fail to enhance deep causal deductive capability. By internalizing temporal causal reasoning capabilities into the model, TCR-Former effectively bridges this gap, providing an efficient and reliable end-to-end foundation for complex reasoning scenarios.

We present TAR (Traffic Anomaly Reasoning), a large-scale multi-task dataset for training and evaluating vision-language models on traffic anomaly understanding. While video anomaly detection (VAD) has traditionally focused on binary classification or temporal localization, TAR moves beyond detection to reasoning. TAR contains 44,040 training annotations with explicit chain-of-thought reasoning traces across 10 task types, covering 3,670 CCTV transportation videos (~26 hours) sourced from eight public datasets. TAR provides a structured progression of tasks organized into three groups — Question Answering, Temporal Reasoning, and Scene Understanding — reflecting our finding that reliable anomaly understanding demands capabilities far beyond detection. Annotations are produced by a hierarchical auto-labeling pipeline using state-of-the-art VLMs, with existing human annotations incorporated as supplementary context for 25% of videos. Every answer is paired with a step-by-step reasoning trace, enabling both supervised training with chain-of-thought supervision and fine-grained evaluation of model reasoning. We evaluate eight vision-language models on a held-out test set of 84 videos (1,008 annotations) with human-reviewed and corrected annotations and find that current models struggle with Temporal Reasoning and Scene Understanding tasks even when they perform well on basic Question Answering — revealing that detection accuracy is a poor proxy for genuine understanding. A multi-task ablation study shows that progressively adding task groups during fine-tuning yields consistent gains, with the full 10-task model achieving +44.5 points on Binary QA accuracy and +21.4 points in aggregate score over the zero-shot baseline, demonstrating the dataset's value as a training resource. TAR is released under CC-BY-4.0 and serves as the official training data for AI City Challenge 2026 Track 3: Anomalous Events in Transportation. The annotations and video download scripts are publicly available at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning.

Human cognition emerges from coordinated spiking dynamics in distributed neural circuits, where information is encoded via both firing rates and precise spike timing determined by brain rhythms. Inspired by this notion, we propose a brain-inspired learning primitive in which cognition-level neural synchrony emerges through iterative *bottom-up* and *top-down* interactions between micro-scale dynamics of spiking neurons and a macro-scale mechanism of oscillatory synchronization. Specifically, we model each parcel (e.g., a cortical region or an image pixel) in the target system as a spiking neuron embedded in a predefined connectivity scaffold. Low-level information is encoded in a spatiotemporal domain, where neurons are selectively grouped and fire spontaneously over time through self-organized dynamics. In the bottom-up route, oscillatory synchronization is formed from past spiking activity accumulated over a finite memory window. Since brain dynamics operate in a regime of partial and transient synchronization rather than global phase locking, we model oscillatory coordination using a time-delayed synchronization formulation, which enables a top-down modulation of heterogeneous neural spiking for a large-scale distributed system. Together, we devise a spiking-by-synchronization neural network (S²-Net) that uses rhythmic timing as a control mechanism for efficient information processing. Promising results have been achieved across a broad range of tasks, including neural activity decoding, energy-efficient signal processing, temporal binding and semantic reasoning.