NeurIPS 2026 论文列表

Large Language Model (LLM) agents are increasingly deployed under structured generation, where grammar constraints enforce output format at inference time. Whereas existing compression methods target unconstrained generation, we first identify a fundamental mismatch: under structured generation, grammar constraints reduce the model's output distribution to valid tokens, and the logits matter most at the small set of positions that decide the agent's action. Motivated by this finding, we propose SKIM (Selective Knowledge-Informed Masking), the first pruning and distillation framework for LLM agents under structured generation, with both stages guided by token role. SKIM partitions token positions into context, decision, and format categories that reflect the input the model reads, the constrained options it chooses among, and the format realized by the constraint mechanism. Specifically, this partition drives two complementary axes: (1) category-aware pruning saliency scores that preserves parameters most relevant for decision positions; and (2) an on-policy distillation objective that aligns student and teacher on valid tokens, paired with category-aware feature matching. We evaluate SKIM on tool-calling benchmarks and within multi-step agent loops, across multiple model scales and sparsity levels. SKIM consistently improves accuracy over baselines, with its superiority becoming most striking in high-sparsity regimes where conventional methods struggle, demonstrating the potential of category-aware compression for deployment-time grammar-constrained agents.

Reinforcement learning from verifiable rewards (RLVR) trains reasoning language models with group-relative objectives such as GRPO. However, it leaves two sources of signal underused: every token in a rollout receives the same response-level advantage, and groups in which all rollouts succeed or all fail have zero group-relative advantage and are typically discarded. We observe that same prompt rollouts are not independent reward labels but sibling attempts: they share scaffolding, diverge at uncertain decisions, and sometimes reconverge. In this paper, we propose Sibling-Relative Advantage Shaping (SIRAS), a lightweight advantage-shaping method that exploits this structure. SIRAS segments rollouts into chunks, aligns sibling chunks with banded dynamic time warping, and reshapes the flat GRPO advantage across tokens via soft-divergence weights that upweight chunks distinguishing a rollout from its siblings and downweight shared ones. This reshaping preserves each rollout’s token-averaged GRPO advantage exactly and reduces to GRPO when weights are constant. The same sibling structure produces calibrated residuals for all-correct and all-wrong groups, reclaiming training signal that group-relative baselines discard. Across six reasoning benchmarks, SIRAS yields the largest gains on AIME-style competition math, including +12.2 on AIME25 with Qwen3-8B and +23.3 on AIME26 in our backbone-scaling study, without extra rollouts, process reward models, or step annotations.

Lifelong cross-modal hashing demands incremental updates as new tasks arrive, yet existing regularization-based methods such as EWC and SI share a fundamental weakness: their effective regularization strength is entangled with task scale, as the empirical Fisher Information Matrix (FIM) scales linearly with sample size. This forces costly per-task hyperparameter tuning in streaming deployments to avoid over- or under-regularizing tasks of varying sizes. We propose \textsc{Sing-CH}, a lifelong hashing framework built on information geometry. At its core is a Gibbs probabilistic reconstruction of the pairwise objective, which gives the parameter space a statistical manifold structure and yields a natural $1/n_t^2$ normalization. We prove this normalization renders the FIM scale-agnostic: its spectral properties remain independent of task size even under diagonal and K-FAC approximations. Beyond this invariance, a Forgetting-Rigidity trade-off theorem shows that natural gradient applied to the composite objective reintroduces scale dependence. This makes our decoupled design, pairing the FIM as regularizer with Adam for optimization, a structural necessity rather than a convenience. Experiments on three benchmarks demonstrate that \textsc{Sing-CH} outperforms 11 baselines by 4.1\% average mAP without per-task tuning. Under extreme task-scale imbalance, degradation is only 2.0\% versus 3.9\% for the strongest baseline, providing the first empirically grounded solution to scale-agnostic lifelong retrieval.

Accurate traffic prediction is essential for urban planning but remains challenging due to the irregular patterns caused by unexpected events. While Spatio-Temporal Graph Neural Networks (STGNNs) effectively model periodic data, they often struggle with these aperiodic fluctuations. To address this, we propose SIMBAD, which improves generalization on irregular patterns via two key mechanisms: 1) adaptive regulation of periodic signals to prioritize recent signals, and 2) dynamic spatial modeling that adjusts influence between connected nodes based on contextual relations. Built upon simple similarity metrics to adapt to aperiodic patterns, SIMBAD outperforms existing methods on real-world benchmark datasets, demonstrating superior performance in predicting unseen and irregular traffic events.

High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry. While effective, this design fragments continuous surfaces into many local tokens, inflates generation cost, and often weakens topological consistency for thin or highly connected shapes. We introduce **SILSA**, a topology-aware 3D generation framework that represents shapes with compact sliding-window slice latents. Instead of generating expensive voxel tokens, SILSA uses a fixed set of overlapping slices along the three canonical axes, where each token summarizes a local depth window to preserve cross-sectional continuity and support single-stage rectified-flow generation. A **Slice VAE** encodes oriented surface samples into multi-axis slice latents and reconstructs them with a sparse volumetric decoder, while a **Volumetric Anchor Lattice** coordinates directional slice streams through a shared 3D workspace. To preserve structural correctness, we introduce slice-level topology supervision that matches persistence diagrams and aligns Betti transitions across neighboring slices. Experiments show that SILSA improves structural fidelity while substantially reducing generation cost. SILSA improves PSNR by 8.7%, coverage by 5.96 absolute points, and Betti error by 9.2% over the strongest baseline, while using 70.0% fewer tokens than the next-most compact baseline and over 98% fewer tokens than sparse or hierarchical tokenizers, reducing training memory by 40.4% and inference time by 58.5%. Qualitative results further show improved preservation of thin structures, repeated components, and long-range connectivity.

Text-driven image editing has advanced rapidly, but reliably localizing these manipulations requires image manipulation localization (IML) models trained on large pixel-annotated datasets, and there is still no low-cost way to obtain such training data at scale. We observe that these data already exist in disguise: public editing datasets contain millions of structurally identical $\textit{(original, edited)}$ pairs to IML training samples, lacking only pixel-level masks. Recovering these masks automatically is non-trivial: pixel differencing is overwhelmed by diffusion-induced perturbations across all pixels, and instruction-only grounding localizes only what the prompt describes, missing unintended editor side-effects. We propose $\textbf{SIGMA}$ ($\textbf{S}$emantic-difference $\textbf{I}$nstruction-$\textbf{G}$rounding $\textbf{M}$ask $\textbf{A}$nnotator), which performs semantic-feature differencing in a vision foundation backbone and injects an instruction-derived spatial prior into this visual stream via bidirectional cross-modal refinement, amplifying the difference signal at intended-edit regions when the editor faithfully realizes user intent. SIGMA is trained in two complementary stages: Stage I supervises on inpainting masks; Stage II closes the diffusion-domain shift via VAE-roundtrip noise calibration, EMA self-training, and an edit-noise disentanglement loss. SIGMA outperforms existing automatic mask generators on five benchmarks ($\textbf{+12.20\%}$ F1, $\textbf{+11.16\%}$ IoU). When applied to public editing corpora, it produces a $\sim$1.1M IML training set that improves six diverse detectors by $\textbf{+18.34\%}$ F1 across five datasets, turning previously unused editing data into a model-agnostic supervisory resource for IML. For reproducibility, our anonymized code is available at https://anonymous.4open.science/r/SIGMA-14FE.

Autoregressive language models generate under a fixed left-to-right order, whereas diffusion language models (dLLMs) denoise masked tokens through flexible any-order trajectories that may commit multiple positions per step. This makes dLLMs a distinctive setting for test-time scaling: additional inference compute can diversify both token choices and the order in which positions are committed. Existing samplers do not explicitly allocate these two sources of diversity. Confidence-based decoding is reliable but often redundant, while global temperature sampling and random remasking inject stochasticity without distinguishing useful exploration from unreliable commitments. We propose SIGMA, a training-free sampler that formulates each denoising step as a constrained allocation problem over decoding efficiency, token-level diversity, trajectory-level diversity, and model-internal commitment uncertainty. The resulting objective decomposes joint sampling entropy into token and trajectory terms, yielding a sigmoid-form stochastic gate for position selection and an adaptive per-position temperature for token sampling. Experiments on LLaDA and Dream across math and code benchmarks show that SIGMA improves the accuracy--cost Pareto frontier and can be integrated into existing dLLM test-time scaling pipelines.

Frontier LLMs are increasingly capable of expert-level scientific reasoning, but using them as reliable scientific agents requires more than general reasoning ability. Scientific simulation is a central test case: modern simulators are configured through executable interfaces such as XML decks, input scripts, and namelists that function as domain-specific languages tied to the simulator's internal API. Translating a researcher's natural-language intent into a runnable configuration remains a recurring expert bottleneck. We study how much of this bottleneck a general-purpose LLM coding harness, Claude Code, can absorb when wrapped with a **Simulator-Interface Grounding Adapter (SIGA)**: a package of skills, tools, and workflow control flow that grounds the agent's outputs in simulator documentation, schema, and example libraries. We instantiate SIGA for GEOS, an open-source multiphysics simulator used in CO₂-storage and induced-seismicity research. A Resolution-IV factorial surfaces three benefits over vanilla Claude Code. First, SIGA improves **reliability**, reducing across-seed variance by roughly 40× by preventing unparseable or empty decks on a hard tail of compound multi-physics tasks. Second, it improves **quality**, raising mean structural similarity by about +7 percentage points on the same hard tail. Third, a self-evolved variant matches the best hand-designed cell with roughly 16% fewer tool calls, suggesting that automatic SIGA discovery is tractable. A preliminary human baseline finds that geoscience-domain-expert volunteers new to GEOS take between 8 and 36 times as long as the agent on a representative task. An explicit human-consultation tool is used in only about 3% of under-specified trials, with the agent instead relying on the on-disk example library as a cheaper retrieval substitute. A small OpenFOAM transfer study indicates that the recipe is not specific to GEOS XML: the same stop-hook component again dominates reliability, and the best SIGA cell outperforms both vanilla Claude Code and a constrained Foam-Agent lint-only baseline. We close with SIGA-design recommendations grounded in failure modes that our best configuration still does not fix.

Multimodal large language models (MLLMs) achieve ever-stronger performance on visual-language tasks. Even as traditional visual question answering benchmarks approach saturation, reliable deployment requires satisfying low error tolerances in real-world out-of-distribution (OOD) scenarios. Precisely, selective prediction aims to improve coverage, i.e.\ the share of inputs the system answers, while adhering to a user-defined risk level. This is typically achieved by assigning a confidence score to each answer and abstaining on those that fall below a certain threshold. To enable reliable generalization, we require reasoner models to produce localized visual evidence while answering, and design a selector that explicitly learns to estimate the quality of the localization provided by the reasoner. We show that SIEVES (Selective Prediction through Visual Evidence Scoring) improves coverage by up to three times on challenging OOD benchmarks (V* Bench, HR-Bench-8k, MME-RealWorld-Lite, VizWiz, and AdVQA), compared to non-grounding baselines. Beyond better generalization to OOD tasks, the design of the SIEVES selector enables transfer to proprietary reasoners without access to their weights or logits, such as o3 and Gemini-3-Pro, providing coverage boosts beyond those attributable to accuracy alone. We highlight that SIEVES generalizes across all five tested OOD datasets and reasoner models (Pixel-Reasoner, o3, and Gemini-3-Pro), without benchmark- or reasoner-specific training or adaptation.

We identify a theoretical incompatibility between 3D point cloud masked autoencoding and SE(3)-equivariant networks. We prove that standard masking triggers a topological collapse, effectively forcing the model to hallucinate orientation without a reference frame. To resolve this, we introduce SIEVE (Scalar Invariant Extraction Via Equivariance), an SE(3)-equivariant self-supervised framework that compresses point cloud geometry into invariant scalar embeddings. At its core is the spherical projection mask, which preserves global pose while obfuscating local semantics, sidestepping the obstruction by construction. Experiments confirm that SIEVE circumvents the predicted collapse where constant and noise masks fail. On downstream point cloud tasks, the resulting embeddings improve over coordinate-based baselines and hand-crafted geometric descriptors, with the largest gains on generalization to unseen shape categories.

Machine unlearning for large language models (LLMs) aims to selectively remove memorized content such as private data, copyrighted text, or hazardous knowledge, without costly full retraining. Most existing methods require a retain set of curated examples to prevent catastrophic degradation of general model utility, creating an extra data dependency that complicates deployment. We propose SHRED (Self-distillation via High-surprisal-only Retain-set-free Entropy Demotion), a retain-set-free unlearning method built on a key insight: not all tokens within a forget set instance carry memorized information equally. High-information tokens concentrate the model's memorized knowledge, while low-information tokens reflect general language competence. SHRED operates in two stages. (1) Selection: We perform a forward pass on a forget set instance, collect per-token autoregressive probabilities, and select the bottom-$P$ (lowest probability, highest Shannon information) as forget positions; the remaining positions are retained as benign anchors. (2) Training: We construct modified KL targets that demote the memorized token's logit at forget positions while preserving the original distribution at benign positions. The model is then trained via a single top-$K$ KL self-distillation objective that simultaneously drives forgetting and utility preservation. We evaluate SHRED across four standard unlearning benchmarks and demonstrate that it establishes a new Pareto-optimal trade-off between forget efficacy and model utility, outperforming retain-set-dependent methods. Our analysis shows that SHRED is robust against relearning attacks and membership-inference attacks, and it maintains stable utility even after many sequential unlearning runs.

Mathematical reasoning benchmarks for large language models face three persistent limitations. First, datasets scraped from public competitions are vulnerable to training-data contamination and saturate quickly. Second, most existing benchmarks cover only part of the olympiad-to-research spectrum. Third, benchmarks that withhold problems indefinitely to mitigate leakage trade away transparency and reproducibility. We introduce \gls{soohak}, a benchmark of 1{,}141 newly authored problems by 86 mathematicians designed to address all three. \gls{soohak} consists of three difficulty splits, constructed adversarially against panels of small, mid-size, and frontier baseline LLMs, spanning olympiad-style problem solving through research-adjacent material; a separate 99-item Refusal split probes recognition of ill-posed prompts. The dataset is temporarily embargoed and evaluated by request, with a committed public release in late 2026. On \splitthree{}, Gemini-3-Pro, GPT-5, and Claude-Opus-4.5 reach Avg@3 of 30.39\%, 26.37\%, and 10.39\%, respectively; the best Pass@3 is only 44.12\%, and 129/340 problems remain unsolved by any closed model evaluated, indicating substantial room for improvement before LMs can meaningfully assist in mathematical research. By contrast, the \calibrationsplits{} are nearly saturated by both closed and open-weight systems. Notably, while GPT-OSS-120B, Kimi-2.5, and GLM-5 each reach Pass@3 between 87.4\% and 88.3\% on \splitone{}, within 6 percentage points of the leading closed model (Gemini-3-Pro, 93.85\%), they trail by over 20 percentage points on \splitthree{}. This pattern suggests open-weight pipelines are well targeted to olympiad-style benchmarks but transfer poorly to research-adjacent material. Finally, through a timed human study with 25 participants of diverse mathematical expertise, we confirm that the difficulty is genuinely mathematical rather than artifactual, with aggregated teams covering 50.6\% of a 79-problem sample.

Graph Neural Networks (GNNs), as representative deep models designed for graph data, effectively learn graph topological information and push the performance boundaries across various graph tasks. However, the computational and memory burden of GNNs poses significant challenges for scaling to large real-world graphs. Spiking Graph Neural Networks (SGNNs), which integrate biologically plausible learning via unique spike-based neurons, have emerged as a promising energy-efficient alternative. Different layers communicate with sparse and binary spikes, which facilitates computation and storage of intermediate graph representations. Despite the proliferation of SGNNs proposed in recent years, there is no systematic benchmark to explore the basic design principles of these brain-inspired networks on the graph data. To bridge this gap, we present **SGNNBench** to quantify progress in the field of SGNNs. Specifically, we elaborately investigate the design space of SGNNs to facilitate the development of a general SGNN paradigm. Regarding efficiency, we empirically compare these baselines w.r.t. model size, memory usage and theoretical energy consumption to reveal previously overlooked energy bottlenecks. Furthermore, SGNNBench establishes a unified and reliable benchmark to comprehensively evaluate 9 state-of-the-art SGNNs across 20 datasets covering diverse graph settings and tasks.

Large language model (LLM)-driven evolutionary frameworks have emerged as a promising paradigm for iterative search in complex optimization and structured generation tasks. However, existing methods typically rely on heuristic variation operators and lack an explicit mechanism to extract directional information from observed performance feedback, limiting both efficiency and reliability. To address this, we propose SGEvolve, a semantic gradient-guided evolutionary framework for LLM-driven optimization. The key idea is to treat the performance differences between parent and offspring as zeroth-order evidence of an underlying ascent direction. Based on this view, we formulate semantic gradient extraction as a Maximum A Posteriori (MAP) estimation problem, where the LLM infers the most probable direction of improvement conditioned on observed evolutionary steps. The inferred gradients are then used to guide mutation and crossover, biasing generation toward more promising regions of the search space. We evaluate SGEvolve on a diverse set of tasks, including mathematical optimization, algorithm design, and real-world industrial problems. Experimental results demonstrate that SGEvolve consistently outperforms existing open-source baselines, achieving superior performance in both solution quality and search efficiency.

Edge of Stability (EoS) refers to the phenomenon where full-batch gradient descent (GD) training of neural networks with step size $\eta$ pushes the largest eigenvalue of the Hessian, i.e. the sharpness, to $2/\eta$ and hovers there. \citet{damian2023selfstab} explains the hovering behavior of the sharpness by *self-stabilization*, a mechanism driven by third-order structure of the loss, and shows that GD implicitly follows projected gradient descent (PGD) on the set where the sharpness is below $2/\eta$. For mini-batch stochastic gradient descent (SGD), the sharpness stabilizes *below* $2/\eta$, with the gap widening as the batch size decreases. However, no theoretical explanation exists for this suppression. In this paper, we introduce *stochastic self-stabilization* that extends the self-stabilization framework to SGD. Our key insight is that gradient noise injects variance into the oscillatory dynamics along the top Hessian eigenvector, strengthening the sharpness-reducing force and shifting the equilibrium below $2/\eta$. Following the approach of \citet{damian2023selfstab}, we define *stochastic predicted dynamics* that tracks the deviation of SGD from the PGD trajectory, and prove a stochastic coupling theorem that relates the SGD sharpness and loss to those of PGD. Based on our predicted dynamic, we derive a closed-form equilibrium sharpness gap that scales with the variance of the gradient noise projected onto the top eigenvector of the Hessian. This formula predicts that smaller batch sizes yield flatter solutions, and recovers GD when the batch equals the full dataset.

Neural networks are known to be susceptible to over-reliance on spurious correlations. However, the precise mechanism by which models exploit shortcut features is not fully understood, and algorithms to mitigate this behavior rely on as yet unjustified assumptions about the learned representations. In this work, we provide the first end-to-end theoretical characterization of spurious feature learning for two-layer ReLU neural networks trained by online minibatch SGD on the logistic loss. We consider data drawn from the high-dimensional Boolean hypercube with a quadratic signal function (namely XOR) and a linear spurious correlation. We show that SGD learns the spurious feature first, and exponentially fast. Moreover, the optimization dynamics couple the spurious and signal features, with a stronger spurious component inhibiting signal feature learning. Our analysis reveals precise phase transitions in the learning dynamics. In the first phase, alignment between the signs of the spurious feature and second-layer weight drives rapid growth of the spurious feature. When the spurious correlation is maximally strong, we show theoretically that the spurious feature dominates even at the sample complexity threshold where XOR would be learned in isolation (i.e., if the spurious feature was absent). In contrast, when the correlation strength is constant, we provide preliminary empirical evidence that the model can eventually learn the XOR signal, although the spurious feature is not forgotten.

Recent mixed-policy optimization methods for LLM reasoning that interleave or blend supervised and reinforcement learning signals report improvements over the standard SFT-then-RL pipeline. We show that numerous recently published research papers rely on a faulty baseline caused by two distinct bugs: a CPU-offloaded optimizer bug in DeepSpeed that silently drops intermediate micro-batches during gradient accumulation (affecting multiple downstream frameworks including TRL, OpenRLHF and Llama-Factory), and a loss aggregation bug in OpenRLHF that incorrectly weights per-mini-batch losses. Together they suppress SFT performance, with the optimizer bug accounting for most of the gap and the loss aggregation bug contributing a smaller additional effect. Once corrected, the standard SFT-then-RL pipeline surpasses every published mixed-policy method we evaluate by +3.8 points on math benchmarks with Qwen2.5-Math-7B and by +22.2 points with Llama-3.1-8B. Even a truncated variant with just 50 RL steps outperforms mixed-policy methods on math benchmarks while using fewer FLOPs.

Cognitive neuroscience research indicates that human spatial navigation relies not merely on abstract global map representations, but rather on anchoring local observations to spatially stable, decision-relevant landmarks to effectively align the current scene with the cognitive map. Inspired by this mechanism, we propose SFPR, a structural fingerprint-based LiDAR-to-OpenStreetMap (OSM) place recognition framework. This framework goes beyond the traditional coarse retrieval paradigm that relies exclusively on global descriptor similarity and further introduces fine-grained discrimination based on structural fingerprints to achieve more accurate place recognition. Specifically, we propose a structural fingerprint-aware attention mechanism that utilizes sparse but spatially stable anchors as saliency prompts, directing the network's attention toward potential fingerprint regions while dynamically generating global descriptors. Subsequently, we formulate the extraction and matching of structural fingerprints as an alternating optimization problem guided by this fingerprint-aware attention. Through iterative optimization, we extract structural fingerprints with high fingerprint-aware attention and geometric consistency from anchors. These structural fingerprints are further fed back into the training stage as geometric priors, effectively suppressing false matches characterized by similar global features but contradictory local structures. Experiments demonstrate that SFPR effectively overcomes the retrieval bottleneck in macroscopically homogeneous scenes, achieving a relative improvement of 30.20\% in Top-1 Recall@1m compared to the current state-of-the-art method. Code are publicly available at https://anonymous.4open.science/r/SFPR.

Vision-Language Models (VLMs) show great potential for Industrial Anomaly Detection (IAD) but often overlook subtle high-frequency defects due to their semantic bias. Additionally, one-class adaptation risks rank collapse and catastrophic forgetting. To address these issues, we propose a unified framework termed SF-DST. First, we design Spatial-Frequency Dual-Stream Transformer, which explicitly captures lost details via a parallel discrete cosine transform branch. Crucially, it incorporates Asymmetric Cross-Modal Modulation (ACMM), a mechanism that leverages spatial semantics as a context gate to selectively retrieve and inject spectral features, effectively compensating for the texture blindness of VLMs. Second, Anomaly-Aware LoRA (A-LoRA) enforces Orthogonal Regularization on the adaptation subspace. This geometric constraint ensures subspace diversity to prevent mode collapse, while optimizing a hypersphere-based metric for precise anomaly discrimination. Extensive experiments on MVTec-AD, VisA and MMAD benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches, successfully bridging the gap between large-model generalization and industrial-grade precision.