NeurIPS 2026 论文列表

Learning constraint-satisfying policies from offline data without risky online interaction is crucial for reliable deployment of reinforcement learning (RL) agents. Representative methods leverage Implicit Q-Learning and advantage-weighted regression to learn value functions and balance the reward–safety trade-off, demonstrating solid safety performance in safety-critical hard constraint scenarios. However, these methods typically couple reward and safety optimization during offline training and rely on a predefined unified factor to balance the reward–safety trade-off (conservatism). Which ignores the fact that different tasks, due to their distinct reward and cost functions, may require different degrees of conservatism to satisfy safety constraints, resulting in sub-optimal performance. To address this limitation, we propose Co3G, a compositional generative model with test-time controllable conservatism for offline safe RL. During training, Co3G decouples the optimization objectives by separately employing reward-optimal guidance and safety-optimal guidance to induce high-reward and high-safety actions. During deployment, Co3G combines the reward and safety objectives via compositional generation, and further provides three mechanisms—manual tuning, rejection sampling, and RL—to balance the reward and safety guidance scales during deployment, thereby achieving a superior reward–safety trade-off. Extensive experiments on the OSRL benchmark show that Co3G delivers strong safety performance under stringent safety constraints while retaining flexible control over conservatism, significantly outperforming previous hard-constraint and soft-constraint baselines.

While autoregressive models optimize the exact likelihood implied by the chain rule, diffusion models are typically trained with denoising objectives. We develop conservation laws based on generalized extrinsic information transfer (GEXIT) functions for a broad class of memoryless noise processes, showing that the data--model cross-entropy (CE) can be characterized \emph{exactly} as an integral of \emph{local} information-theoretic derivatives along the noise path. This yields a unified characterization of the likelihood for discrete diffusion and continuous diffusion, with the Gaussian case reducing to the well-known I-MMSE relationship connecting mutual information and estimation theory. An immediate implication is a \emph{locality} property: one can compute the information-theoretic derivatives using only the marginal posteriors along the noise path. As a result, training reduces to learning the marginal posteriors by minimizing the negative log-likelihood. While the conservation law implies that the entropy does not depend on the noise path, finite-capacity denoisers approximate the posteriors with varying accuracy across noise types, leading to differences in performance. We validate these predictions on synthetic Markov sources and standard benchmarks, including \texttt{text8} and CIFAR-10.

A local specialist LLM, fine-tuned with reinforcement learning from verifiable rewards (RLVR) on operator-local data, is installed inside a single regulated organization under a per-deployment error budget $\alpha$. The operator needs a safety certificate that holds on *this* deployment's stream, simultaneously at every wall-clock round: no pooling across deployments, no waiting for a long-run average. Existing wrappers cannot deliver this on these adaptive, online-updated streams: offline conformal-risk methods require exchangeability; online-conformal methods bound only long-run averages; non-exchangeable extensions are marginally valid; and the closest anytime wrapper, A-RCPS, controls marginal rather than selective risk. Through a (test statistic, validity guarantee, deployment rule) framework we identify one empty cell forced by the deployment requirements (e-process per threshold, selective risk, anytime-pathwise validity, max-certified-threshold rule); **Conformal Selective Acting** (CSA) fills it as a per-round wrapper maintaining a Ville-type e-process per candidate score threshold on a Bonferroni grid, evaluated against the RLVR filtration. Under predictable updates and isotonic-calibrated monotone risk we prove (i) an anytime-pathwise selective-risk bound $R_T^{\mathrm{act}} \le \alpha + O(N_T^{-1/2})$, (ii) rate-optimal certification matching $\Theta(\bar\eta^{-2} \log(1/\delta))$, and (iii) a horizon-independent release-rate gap. Across eight specialist benchmarks (480 streams), sixteen adversarial distribution-shift cells (160 streams), and five live Expert-Iteration RLVR cells with online LoRA over four base models in three architecture families (10,300 rounds), CSA is the only method among ten directly compared that satisfies both pathwise validity and non-refusing deployment on every cell. We do not propose a new LLM, training algorithm, or policy class; CSA is the deployment-side complement, orthogonal to the model itself, for operators who cannot use a frontier API.

Deep-learning protein structure predictors achieve near-experimental accuracy on individual folds, yet their default inference samples concentrate around a single dominant conformation. We introduce ConforFlux, an inference-time procedure for Boltz-2 that couples $M$ structure-prediction trajectories through a pairwise C$\alpha$-RMSD repulsion gradient on the trunk's single and pair embeddings. Because the trunk conditions every block of the diffusion module, this update propagates to every subsequent denoising step. On four conformational-change categories, ConforFlux improves the per-state success rate over Default Boltz-2 by 3–17 percentage points while preserving Default-comparable physical quality. On twelve transporter pairs with at least one alternate-state reference released after the Boltz-2 cutoff, the $2$ Å success rate rises from $4/12$ to $9/12$. On the human dopamine transporter, ConforFlux additionally reaches the post-cutoff inward state, which $500$ Default samples never attain.

MLLM-based detection has shown promising end-to-end detection ability in recent years, but its generated detection results still lack reliable confidence for evaluation and interpretation. Since detection data itself do not provide human-annotated confidence labels, existing MLLM-based detection methods mostly obtain confidence implicitly from prompted self-assessment, generation probabilities, or external proxy scores. While these scores are available for evaluation, they are not explicitly connected with the actual detection quality. Inspired by conventional detectors that optimize confidence through classification, objectness, or quality estimation branches, we formulate confidence in MLLM-based detection as an explicit modeling target to be learned, optimized and calibrated. Specifically, we present ConfDet, a practical confidence framework that first uses discrete confidence tokens based supervised fine-tuning to enable stable and parsable confidence generation, then applies GRPO with detection and separation joint rewards to optimize confidence score separation and preserve detection quality, and finally performs condition-aware post-hoc calibration to align confidence with empirical accuracy. Experiments across different in-distribution and out-of-distribution benchmarks show that, ConfDet greatly improves AP through better confidence ranking and reduces confidence miscalibration across benchmarks, providing a reliable confidence for MLLM-based detection.

Multimodal Large Language Models (MLLMs) trained on sequential tasks suffer from catastrophic forgetting. We study this problem in LoRA-based multimodal continual instruction tuning. By analyzing the diagonal Fisher information of LoRA parameters, we find that forgetting is amplified not merely by shared high-sensitivity parameters across tasks, but by the concentration of updates on those shared parameters, \ie, some tasks distribute their change uniformly across parameters, while others pack most of it into a narrow subset, causing disproportionate interference on the parameters that prior tasks depend on. We further observe that this concentration manifests in the dominant directions of recent gradients. Based on this insight, we propose \textbf{D}om\textbf{i}nant-direction \textbf{G}radient \textbf{Pro}jection (\textbf{DiGPro}), which maintains a short buffer of recent gradient directions, extracts their dominant subspace with adaptive rank selection, and attenuates the parallel component for each optimizer step. Crucially, DiGPro requires no prior-task data or stored subspaces, operating on the current gradient history. Experiments on cross-modal and vision-centric continual learning benchmarks show DiGPro consistently reduces forgetting and attains competitive final results. The code is available in the supplementary materials.

Large language models (LLMs) are costly to deploy due to their large memory footprint and high inference cost. Weight-activation quantization can reduce these costs, but low-bit activation quantization remains difficult because activation outliers induce large quantization error. Recent rotation-based methods address this by applying orthogonal transformations that redistribute activation magnitude across dimensions, but existing approaches either require expensive end-to-end rotation training or rely on stored activation corpora, introducing significant compute or storage overhead. We propose a lightweight post-training rotation calibration method for LLM activation quantization. Our method learns orthogonal rotations that align normalized activations with the corners of an inscribed hypercube, encouraging activation energy to be distributed more evenly across dimensions. This objective admits an efficient closed-form update via the orthogonal Procrustes problem, avoiding gradient-based optimization over the orthogonal group. We further introduce an online calibration procedure that updates rotations as calibration samples are processed, eliminating the need to store activations on disk and allowing rotations to adapt to quantized activation distributions during calibration. Experiments on Llama-2 and Llama-3 models from 3B to 70B parameters show that our method achieves competitive or improved performance across perplexity benchmarks and common sense reasoning tasks while avoiding both costly end-to-end training and large offline activation storage.

Competition among LLM providers hinges not only on model accuracy, but also on the latency of serving user requests. However, user demand leads to congestion which increases latency, creating a feedback loop between provider decisions and user choices. In this work, we study a stylized game between two providers who strategically balance allocating fixed compute between training better models and reducing latency. A population of users then chooses between the providers. Our equilibrium analysis reveals that a lower-budget provider can still attract users by taking advantage of the congestion faced by the dominant provider. Relative to a competitive baseline where providers have equal compute resources, compute asymmetry leads the higher-resource provider to divert a greater fraction of their compute from improving accuracy to reducing latency. Moreover, this compute asymmetry strictly harms users. Altogether, our results illustrate how congestion distorts competition between model-providers, leading to subtle implications for market competitiveness, model accuracy, and user utility.

Learning Gibbs distributions using only sufficient statistics has long been recognized as a computationally hard problem. On the other hand, computationally efficient algorithms for learning Gibbs distributions rely on access to full sample configurations generated from the model. For many systems of interest that arise in physical contexts, expecting a full sample to be observed is not practical, and hence it is important to look for computationally efficient methods that solve the learning problem with access to only a limited set of statistics. We examine the trade-offs between the power of computation and observation within this scenario, employing the Ising model as a paradigmatic example. We demonstrate that it is feasible to reconstruct the model parameters for a model with $\ell_1$ width $\gamma$ by observing statistics up to an order of $O(\gamma)$. This approach allows us to infer the model's structure and also learn its couplings and magnetic fields. We also discuss a setting where prior information about structure of the model is available and show that the learning problem can be solved efficiently with even more limited observational power.

We study the degree-weighted work required to compute $\ell_1$-regularized PageRank using the standard one-gradient-per-iteration accelerated proximal-gradient method (FISTA). For non-accelerated local methods (ISTA), the best known worst-case work scales as $\widetilde{O} ((\alpha\rho)^{-1})$, where $\alpha$ is the teleportation parameter and $\rho$ is the $\ell_1$-regularization parameter. A natural question is whether FISTA can improve the dependence on $\alpha$ from $1/\alpha$ to $1/\sqrt{\alpha}$ while preserving the $1/\rho$ locality scaling. The challenge is that acceleration can break locality by transiently activating nodes that are zero at optimality, thereby increasing the cost of gradient evaluations. We analyze FISTA with respect to a less regularized objective and show that, under a checkable confinement condition, all spurious activations remain inside a boundary set $\mathcal{B}$. This yields a bound consisting of an accelerated $(\rho\sqrt{\alpha})^{-1}\log(\alpha/\varepsilon)$ term plus a boundary overhead $\sqrt{vol(\mathcal{B})}/(\rho\alpha^{3/2})$. We provide graph-structural conditions that imply such confinement. Experiments on synthetic and real graphs show the resulting speedup and slowdown regimes under the degree-weighted work model.

Reconstructing the motion of objects from videos is a key component for embodied AI and robot manipulation. While diverse approaches to object pose tracking have been studied, they rely heavily on strong external priors, such as depth data or 3D templates, and remain highly vulnerable to severe occlusions by hand grasps despite the use of explicit masks. In this work, we present ComPose, a 6DoF object tracking framework designed for hand-aware object pose estimation from RGB video. Rather than treating the hand purely as an occluder, our method harmonizes hand motions as a complementary cue for object tracking. In detail, we recover a variety of object motions over time by combining object- and hand-derived cues from foundation models within a unified tracking pipeline. Here, ComPose adaptively selects informative hand joints, blends object- and hand-based rotation cues, and refines translation from the visible geometric evidence using a learned correction. We further enforce the temporal consistency over both rotation and translation, yielding stable 3D object trajectories over time without any external smoothing. Extensive experiments show that our method is accurate, efficient, and robust under severe hand occlusion and geometric ambiguity. In addition, the resulting trajectories can also effectively transfer to downstream robot manipulation by enabling robots to reconstruct human actions from online videos.

Self-improvement is a critical capability for large language models and other intelligent systems, enabling them to refine their behavior and internal consistency without external supervision. Despite its importance, prior approaches largely rely on empirical heuristics and lack formal guarantees. In this paper, we propose a principled framework for self-improvement based on the concept of coherence, which requires that a model's outputs remain consistent under task-preserving transformations of the input. We formalize this concept using projection-based mechanisms that update a baseline model to be coherent while remaining as close as possible to its original behavior. We provide rigorous theoretical guarantees that these mechanisms achieve monotonic improvement, measured by a reduction in expected Bregman divergence. Our analysis is comprehensive, covering both \emph{direct} and \emph{two-step} projection methods, and robustly extends these guarantees to non-realizable settings, empirical (finite-sample) distributions, and relaxed coherence constraints.

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.

Neural networks are expressive predictors capable of modeling rich, multi-modal outcome distributions. However, this expressive power comes with challenges in ensuring robust predictions—a critical requirement in safety-critical domains. Randomized smoothing is a leading technique for improving robustness, particularly against adversarial perturbations. Yet, in stochastic multi-modal regression settings, randomized smoothing often fails due to mode collapse, yielding averaged predictions that do not reflect the underlying distribution. To address this limitation, we propose clustered $\alpha$-smoothing, a framework that (1) partitions noisy samples using an arbitrary clustering algorithm, (2) applies $\alpha$-smoothing locally within each cluster, and (3) combines the resulting predictions into a mixture distribution. By interpreting the smoothing distribution as a mixture of $\alpha$-smoothers, we derive a lower bound on the probability that the smoothed prediction lies within a union of compact regions corresponding to distinct modes. We empirically evaluate our framework on two benchmarks, demonstrating substantial improvements over state-of-the-art methods. In stochastic trajectory prediction on a driving simulator dataset, our approach achieves, on average, a 33\% lower Wasserstein distance to the ground-truth distribution compared to $\alpha$-smoothing. In quadrotor control, where modes correspond to distinct feasible paths to a target, our method reduces the collision rate by 81\% relative to the state-of-the-art randomized smoothing.

Large-scale approximate nearest neighbor search commonly relies on partitions for indexing: database vectors are partitioned into clusters, and for each query a probing function selects the clusters to be scanned. The query probing function and the database partition are rarely treated as separate entities: most techniques assign queries with the same assignment function as the database vectors, which is suboptimal especially when database and query distributions differ. This paper introduces CwA (Cluster with Auctions), which addresses this limitation by jointly learning a balanced database partition and a neural probing function. CwA optimizes search performance directly for the query distribution. It minimizes its objective by alternating two steps: (i) gradient descent on the neural network of the probing function, and (ii) a large-scale combinatorial optimization of the cluster assignment for the database vectors. We solve the latter with a parallelizable auction algorithm that balances the partition by design. To further scale CwA, we extend the method to a Cartesian product of clusters that increases the partition’s granularity. When database and query distributions differ, CwA achieves up to 4.7$\times$ throughput over the state of the art at equal recall. In the in-distribution (ID) setting, even a simple linear probing function trained with CwA outperforms competing deep neural methods.

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.

Unified and scalable Transformers have recently achieved remarkable success in modeling diverse phenomena traditionally associated with computer graphics, such as 3D visual effects, rendering processes, and motion in videos. In this work, we take a step further by investigating whether modern Transformer techniques can tackle the challenging task of cloth simulation. To this end, we present ClothTransformer, a framework that reformulates cloth simulation as autoregressive sequence modeling in a learned latent space. Existing neural cloth simulators are largely specialized to single scenarios, intrinsically coupled to the mesh discretization, and lack robust collision handling. Our approach addresses these limitations through three contributions: (1) a unified Transformer architecture that handles diverse scenarios---body-driven garments, robotic manipulation, and free-fall collisions---under a single model and achieves approximately $4$--$9{\times}$ lower error than prior state-of-the-art across all scenarios; (2) a scalable latent-space formulation that compresses arbitrary-resolution meshes into a fixed-size set of latent tokens, making temporal dynamics computation independent of mesh resolution; and (3) a diverse-scenario high-fidelity penetration-free dataset of ${\sim}$493.4k frames spanning all three settings, which enables a differentiable Continuous Collision Detection (CCD) module to suppress penetration artifacts.

In a standard transformer, each layer computes attention once: queries examine keys, select values, and move on. But attention is a function of representation, and representation is a function of attention. We close this loop. Fixed-Point Self-Attention (FPSA) computes its attention pattern by iterating the query/key/value process to convergence. Values $V$ are derived from the layer input and held fixed; queries and keys are recomputed from an evolving output state $Z_t$ at each step, converging when representations reach equilibrium, $Z^* = f(Z^*)$. FPSA is a drop-in replacement for multi-head attention that adds zero parameters and adapts depth on its own: easy tokens stabilize in $2$--$3$ iterations, hard tokens refine over $15$ or more, with $\mathcal{O}(1)$ memory cost regardless of the number of iterations. It lifts BERT-Base and ELECTRA-Base on GLUE and SQuAD~v2.0, improves ViT-B/16 accuracy by up to 20\%, and delivers matching gains on vision-language tasks, all without adding a single parameter. Unlike looped transformers that iterate entire layers, FPSA iterates only the attention step, so overhead is modest: a median of $3$--$6$ steps per layer adds roughly $1.6\times$ GFLOPs and $1.3$--$1.4\times$ wall-clock time over BERT-Base. On multi-step reasoning benchmarks (GSM8K, BBH, LogiQA), FPSA's adaptive computation yields clear gains, letting the model refine longer when the input demands it.

Recovering dynamical systems from noisy observations is a recurring challenge across scientific domains, including neuroscience and physics. Latent stochastic differential equations (SDEs) address this by modeling the system as an unobserved state that evolves according to a learnable SDE and generates the observations. Variational inference (VI) provides a tractable objective for fitting latent SDEs. Traditional VI algorithms evaluate this objective by numerical simulation over a time discretization, trading fidelity for computational cost. A recent class of algorithms, simulation-free VI, sidesteps this tradeoff by parameterizing the posterior through its instantaneous marginals rather than its drift. In this work, we show that the efficiency of simulation-free VI algorithms comes at a price: their parameterizations restrict the approximate posterior to a subset of the SDEs available to simulation-based methods, degrading both posterior inference and parameter learning. We propose Helmholtz-SDE, a simulation-free VI algorithm that closes this gap by optimizing over path laws compatible with a prescribed collection of marginals. Helmholtz-SDE recovers dynamics more faithfully than prior simulation-free methods, with the largest gains under high posterior uncertainty. It further matches the performance of simulation-based VI at a fraction of the runtime.