NeurIPS 2026 论文列表

Exploration has been a focus of reinforcement learning research for a long time. Recently, there has been growing evidence that encouraging exploration during LLM reinforcement learning can improve downstream performance. However, methods for controlling exploration often rely on heuristics like clipping or temperature scaling that are unable to fundamentally change token-level distributions, which might limit exploration. Here, we explore controlling exploration via variational learning, where a distribution over neural network parameters is learned via noisy optimization, with parameters sampled from an approximate posterior. We introduce Perturbed Parameter Policy Optimization (3PO), where the amount of noise that is added to model parameters functions as an additional control lever that can be tuned for exploration. We show that using multiple parameter samples within a batch performs better than using a single sample. When using multiple parameter samples in group-based methods like GRPO, we find that calculating advantages across rollouts from all parameter samples performs best. We call this Chunked Perturbed Parameter Policy Optimization (C3PO). We show empirically on a range of math reasoning benchmarks and LLMs that using multiple parameter samples can improve downstream performance, especially on harder benchmarks like AIME. Overall, our work presents evidence that parameter-space exploration can improve LLM reinforcement learning.

Continuous-token autoregressive (AR) generation directly models continuous signals without discretizing them into vocabulary tokens, but suffers from a severe train--inference mismatch: teacher-forced training conditions on ground-truth prefixes, whereas inference conditions on model-generated continuous tokens. This mismatch becomes especially severe in pixel-space image generation, where each token is a high-dimensional raw pixel patch and autoregressive errors can accumulate over generation steps. Rollout-based training can reduce this mismatch but is prohibitively slow, especially when each token is produced by a multi-step diffusion head. We propose \emph{Parallel Rollout Approximation} (PRA), which approximates rollout-based training by constructing training inputs in parallel at all positions. PRA combines a shared one-step denoiser with an end-to-end learned low-dimensional intermediate state, aligning these constructed training inputs with inference-time generated outputs without relying on a separately pretrained tokenizer. On class-conditional ImageNet-1K generation at 256$\times$256 resolution, even PRA-S with 135M parameters achieves an FID of 2.58, surpassing the previous billion-scale pixel-space AR result of 3.60. Scaling to PRA-L (511M) further improves FID to 1.94, establishing a new state of the art among pixel-space AR models and substantially narrowing the gap to pixel-space diffusion models. Beyond generation, PRA achieves higher ImageNet classification probing accuracy than latent-space AR baselines, suggesting its potential for unified pixel-space image generation and understanding.

Entropy-regularized optimization problems arise in various areas of machine learning, including mean-field neural networks, variational inference, and reinforcement learning. Mean-field Langevin dynamics (MFLD) is a dynamics capable of sampling from the optimal solution of such problems under appropriate conditions, and has been the subject of extensive recent research. As machine learning models grow in scale and parameter dimension, computational efficiency becomes increasingly important, especially in reducing the adaptive complexity of computation of MFLD, that is, the number of sequential rounds they require. In this work, we initiate the study of the adaptive complexity of MFLD and provide the first theoretical evidence that MFLD can indeed be accelerated through parallelism. In particular, we provide polylogarithmic convergence guarantees for both the infinite-particle and finite-particle setting under the log-Sobolev inequality.

Training large multimodal models (LMMs) via reinforcement learning (RL) to natively invoke video-processing tools (e.g., cropping) has become a promising route to long-video understanding. However, existing native-RL methods dispatch tool calls sequentially (i.e., one per turn): a single wrong crop propagates errors without peer correction, multi-turn tool calls corrupt context, and inference cost scales linearly with the number of turns. We introduce **ParaVT**, the first multi-agent end-to-end RL-trained framework for **Para**llel **V**ideo **T**ool calling, dispatching multiple time-window crops in a single turn for cleaner context and better fault tolerance. Yet applying standard RL to ParaVT reveals an obstacle we term the *Tool Prior Paradox*: the pretrained tool priors that enable tool exploration also destabilize cold-started structural format and expose the skip-tool reward shortcut under temperature sampling. A cross-model contrast on a weaker-prior LMM supports this claim: format stays stable but RL elicits zero tool calls, indicating that prior strength is the shared driver of both format collapse and tool exploration. We propose **PARA-GRPO** (**P**arseability-**A**nchored and **R**atio-g**A**ted **GRPO**), which augments standard RL with two complementary mechanisms: (i) a *targeted format reward* applied only at the structural-token positions most prone to collapse, and (ii) a *per-prompt frame-budget randomization* that creates training prompts where calling the tool yields a measurable reward signal over skipping it. Across six long-video understanding benchmarks, ParaVT improves over the Qwen3-VL baseline by $+7.9\%$ on average, with PARA-GRPO lifting training-time format compliance from $0.13$ to $0.64$. As tool capabilities become increasingly internalized in modern LMMs, RL must cooperate with the resulting priors, and ParaVT offers a general recipe for agentic RL. Code, data, and model weights will be released.

Flow matching and diffusion models achieve high-quality generation by solving continuous-time generative dynamics, but their sampling remains inherently sequential and costly. Parallel sampling reduces wall-clock latency, yet often requires many function evaluations (NFEs) due to inaccurate future-state initialization and repeated correction. We propose ParaPC-FM, a fixed-point predictor--corrector framework that unifies high-order solver correction and future-state initialization, enabling both components to be incorporated into different parallel samplers. In particular, our UniP-based truncated initialization reuses past velocity predictions to estimate future states more accurately, reducing correction iterations and NFEs with almost no additional overhead. On a ParaDiGMS-style framework, ParaPC-FM achieves up to $1.25\times$ and $1.49\times$ speedups on FLUX.1-dev at $50$ and $100$ steps, and up to $1.22\times$ and $1.16\times$ speedups on Stable Diffusion 3 and Stable Diffusion 3.5 at $50$ steps. When integrated into ParaTAA, the proposed initialization further achieves up to $1.17\times$ speedup at $50$ steps. Importantly, these acceleration gains are obtained while keeping image quality nearly unchanged.

Peer review is sometimes seen as noisy and difficult to predict. How much of the final accept/reject decision is recoverable from the paper itself? We investigate this question with PaperLens, text and vision models trained to predict binary acceptance from anonymized paper artifacts. For a clean evaluation, we construct balanced, shortcut-controlled OpenReview and source-derived arXiv datasets that remove deanonymization artifacts and majority-class shortcuts. Supervised binary decision training is surprisingly strong. PaperLens outperforms frontier prompting and review-trained baselines on decision accuracy and ranking, while its acceptance probability remains well calibrated after validation-set scaling and correlates better to ratings than baselines. Vision consistently improves over markdown text, showing that layout, figures, and visual presentation carry reviewer-relevant signal. Scaling from 3B to 14B parameters yields only modest gains, with clean supervision and faithful paper representations mattering more than model size. Together, our results show that decision prediction provides a powerful signal for paper assessment and can even improve the alignment of AI reviewing agents.

Modern text-to-image models produce high-fidelity images but still struggle with compositional prompts that require instance identity, attribute ownership, counting, spatial ordering, and role-sensitive relations. We introduce *Panoptic Scene Program Diffusion Transformer* (PSP-DiT), a diffusion-transformer architecture that treats a panoptic scene program as a first-class latent variable rather than an external control signal or post-hoc parse. PSP-DiT jointly denoises image latents and scene-program latents through coupled transformer streams, while panoptic grounding and cycle-consistency objectives tie object instances, attributes, relations, and counts to visual support in the generated image. Under matched training and inference settings, PSP-DiT improves over a strong flat-text baseline across GenEval 2, SANEval-Simple, PSG-Score, and DetailMaster, with the largest gains on counting, attribute binding, role-sensitive relations, and long structured prompts. The method preserves image quality, adds modest inference overhead, and remains robust to imperfect scene programs.

We introduce Panoptic Saliency Ranking (PSR), a new task that aims to estimate the visual saliency of all objects within a scene. Unlike salient object ranking, which focuses only on ranking salient regions, PSR extends the problem to the entire scene, providing a holistic and fine-grained understanding of visual saliency. To facilitate this task, we construct PSR18K, the first large-scale benchmark for panoptic saliency ranking. It contains 17,861 images with 146,460 annotated object instances spanning 13 superclasses and 376 fine-grained categories. In addition, PSR18K provides relation graph annotations to explicitly model semantic and spatial relationships among objects. We further propose an efficient Intrinsic Relation Graph Network (IRGNet), which reinterprets Transformer self-attention as an implicit relational matrix and seamlessly transforms it into an explicit relation graph, thereby obviating the need for additional relation modeling and significantly improving inference speed. Extensive experiments on PSR18K and three widely used SOR benchmarks demonstrate that our IRGNet achieves new state-of-the-art performance and faster inference with fewer parameters. The dataset and code are publicly available at https://sites.google.com/view/PSR18K.

Multimodal large language models (MLLMs) still struggle with spatial understanding under the dominant perspective-image paradigm, which inherits the narrow field of view of human-like perception. For navigation, robotic search, and 3D scene understanding, 360$^\circ$ panoramic sensing offers a form of supersensing by capturing the entire surrounding environment at once. However, existing MLLM pipelines typically decompose panoramas into multiple perspective views, leaving the spherical structure of equirectangular projection (ERP) largely implicit. In this paper, we study pano-native understanding, which requires an MLLM to reason over an ERP panorama as a continuous, observer-centered space. To this end, we first define the key abilities for pano-native understanding, including semantic anchoring, spherical localization, reference-frame transformation, and depth-aware 3D spatial reasoning. We then build a large-scale metadata construction pipeline that converts mixed-source ERP panoramas into geometry-aware, language-grounded, and depth-aware supervision, and instantiate these signals as capability-aligned instruction tuning data. On the model side, we introduce PanoWorld with Spherical Spatial Cross-Attention, which injects spherical geometry into the visual stream. We further construct PanoSpace-Bench, a diagnostic benchmark for evaluating ERP-native spatial reasoning. Experiments show that PanoWorld substantially outperforms both proprietary and open-source baselines on PanoSpace-Bench, H$^\ast$Bench, and R2R-CE Val-Unseen benchmarks. These results demonstrate that robust panoramic reasoning requires dedicated pano-native supervision and geometry-aware model adaptation. All source code and proposed data will be publicly released.

We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications.

Heterogeneous AI systems composed of multiple models, architectures, harnesses, or inference-time settings can improve quality and efficiency by routing queries to the \emph{specialist} who can answer most effectively at the lowest cost. Routing requires estimating each specialist's expected return, but this value estimation has a cost. Cheap estimators (e.g., embedding-based predictors) are fast but noisy, while accurate estimators (e.g., fine-tuned models with access to retrieval results or partial reasoning traces) are expensive. We formalize this tradeoff as an instance of Pandora's Box, the classical problem of optimal search with costly inspection. Under a Gaussian signal model, the resulting policies have closed-form value-of-information expressions that determine, for each specialist and input, whether refining the value estimate is worth its cost. We call the centralized policy Pandora's Router. We extend this to a decentralized setting, Pandora's Bidder, where specialists independently decide whether to invest in self-assessment before accepting an offered price to claim a query. Experiments across three domains---a standard multi-LLM benchmark, retrieval-augmented specialists, and LLMs with variable inference-time reasoning---show that Pandora's Router matches the routing quality of exhaustive estimation, while querying the expensive estimator far less often. In the decentralized setting, value-of-information reasoning improves allocative efficiency when competing estimates are accurate; when competing estimates are noisy, however, it can increase the strategic specialist's utility at the expense of

Foundation models (FMs) have shown great promise in medical imaging, but most FMs are trained on unimodal data within isolated domains, such as brain MRI alone. Human aging and disease arise through coordinated biological processes across organs, therefore motivating multimodal FMs that learn whole-body representations. A key challenge, however, is that real-world multimodal biomedical data are often missing not at random, which can reduce power, limit generalizability, and introduce bias. We propose Pan-FM, a pan-organ foundation model pre-trained on imaging from seven organs (Brain, Heart, Adipose, Liver, Kidney, Spleen, and Pancreas) under realistic missing-organ scenarios. Pan-FM uses a unified backbone that handles organ missingness during both training and inference, and is pre-trained with masking-based self-distillation. We find that naive multimodal pre-training leads to dominant-organ shortcut learning bias, with the model over-relying on dominant organs such as adipose and heart. To address this, we introduce Saliency-Guided Masking (SGM), which uses the model attention distribution to adaptively mask dominant organs during pre-training, thus encouraging more balanced cross-organ, whole-body learning. Notably, SGM introduces negligible computational overhead and can be seamlessly integrated into existing self-supervised learning frameworks to improve multi-organ representation learning. On the UK Biobank, Pan-FM achieves stronger prediction across 13 disease categories and 14 single disease entities than single-organ and multi-organ baselines, with improved robustness under missing-organ settings. Pan-FM serves as a scalable solution to realistic modality-missingness in multimodal learning in system neuroscience and as a step toward more generalizable whole-body FMs.

Explicitly modeling periodicity has proven to be an efficient approach for time series forecasting, where the residual, obtained after removing the periodic component from the time series, is typically treated as unstructured noise. However, in this work, we show that such residuals still exhibit structured, \textit{phase-dependent} deviations, as supported by both empirical observations and theoretical analysis. This suggests that residuals should not be treated merely as unstructured noise, but instead as \textit{phase-conditioned signals} whose variations arise in both direction and magnitude. Motivated by this perspective and inspired by phase and amplitude modulation in classical signal processing, we propose \textit{Paloma}, a simple yet effective framework for modeling residual deviations. Specifically, Paloma comprises two complementary modules: \textit{a Phase Rotation module}, which performs phase-conditioned rotations to capture directional changes in the feature space, and \textit{an Amplitude Scaling module}, which applies phase-conditioned affine transformations to model magnitude variations. We further abstract the two modules into a general-purpose plugin, termed the \textit{Paloma technique}, which can be readily integrated into existing forecasting architectures to modulate input sequences in a phase-aware manner, enabling backbone models to capture phase-dependent variations in both direction and magnitude. Extensive experiments on twelve datasets demonstrate that Paloma achieves state-of-the-art and robust forecasting performance. Additional results show that the Paloma technique consistently improves a wide range of baseline models, validating its effectiveness as a general-purpose component.

Current safety alignment of foundation models largely follows a \emph{one-size-fits-all} paradigm, applying the same refusal policy across users and contexts. As a result, models may refuse requests that are unsafe for general users but legitimate for authorized professionals, limiting helpfulness in specialized professional settings. Existing approaches either require costly realignment or rely on inference-time steering that suffers from imprecise control and added latency. To this end, we propose \textsc{Palette}, a modular, controllable, and efficient framework that selectively relaxes refusal behavior on authorized target domains while preserving standard safety elsewhere. Our method identifies a refusal direction via multi-objective search and internalizes it into the model through lightweight adaptation. \textsc{Palette} further supports modular composition: it learns domain-specific safety controls independently and composes them through parameter merging, enabling on-demand multi-domain authorization without retraining. Experiments across four safety benchmarks, multiple model variants, and both LLMs and VLMs show that \textsc{Palette} delivers precise safety control without sacrificing general utility, offering a practical path toward foundation models that adapt to diverse professional needs.

The evolution of colour vision is captivating, as it reveals the adaptive strategies of extinct species while simulta- neously inspiring innovations in modern imaging technol- ogy. In this study, we present a simplified model of vi- sual transduction in the retina, introducing a novel opsin layer. We quantify evolutionary pressures by measuring ma- chine vision recognition accuracy on colour images shaped by specific opsins. Building on this, we develop an evolu- tionary conservation optimisation algorithm to reconstruct the spectral sensitivity of opsins, enabling mutation-driven adaptations to to more effectively spot fruits or predators. This model condenses millions of years of evolution within seconds on GPU, providing an experimental framework to test long-standing hypotheses in evolutionary biology , such as vision of early mammals, primate trichromacy from gene duplication, retention of colour blindness, blue-shift of fish rod and multiple rod opsins with bioluminescence. More- over, the model enables speculative explorations of hypo- thetical species, such as organisms with eyes adapted to the conditions on Mars. Our findings suggest a minimalist yet effective approach to task-specific camera filter design, op- timising the spectral response function to meet application- driven demands. The code will be made publicly available upon acceptance.

Direct pairwise AUC-surrogate training is attractive because it trains toward the ranking metric used at test time, especially under class imbalance. This creates a direct alignment between the training objective and the evaluation metric, but such metric alignment alone does not guarantee that stochastic training remains corrective. We study this gap through Corrective Power: whether pairwise mini-batch updates produce stronger and more reliable improvements for more severe positive-negative ranking errors. Our main observation is that direct pairwise training can enter regions where severe errors remain common but no longer produce useful corrective motion. In such regions, even the mini-batch noise that would normally help SGD move along the steepest severe-error correction direction can become much weaker. In this paper, we seek a theoretical explanation for this by exploring the degenerate property of the U-statistics formed from the pairwise gradients therein. Our results provide a mechanistic explanation for several empirical patterns: (a) squared-hinge loss is often an effective AUC surrogate; (b) cross-entropy warm-up can help models that enter the AUC phase from a poor starting region; and (c) parameter-efficient fine-tuning from a pretrained model can reduce that need by starting closer to a favorable ranking geometry. Experiments on 8 CV tasks and 8 NLP tasks, using ResNet50, DenseNet121, DistilBERT-base, and Qwen2.5-1.5B-Instruct, support these predictions.

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the \emph{branch externality} and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce \textsc{Pace}, a per-step admission controller that treats extra branches as opportunistic work, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, \textsc{Pace} improves goodput by $1.77\times$ over \textsc{IRP-Off} and by $1.48\times$ over \textsc{IRP-Eager}, while maintaining over 95\% SLO attainment.

Existing human-activity datasets typically cover only one or two sensor channels and often focus on short, isolated actions, leaving long-duration compositional tasks largely unsupported as benchmarks for temporal cross-modal sensorimotor learning. We introduce PULSE ($\mathbf{P}$hysiological $\mathbf{U}$nified $\mathbf{L}$ong-duration $\mathbf{S}$ynchronized $\mathbf{E}$mbodiment), a synchronized five-modality dataset that enables benchmarking temporal cross-modal sensorimotor learning: how models recognize ongoing activities, anticipate future contact, reconstruct unobserved signals, and remain robust when physiological, motion, gaze, inertial, and tactile channels are partially observed. Collected from 40 volunteers across 8 ecologically valid daily-activity scenarios together with a dedicated motion-primitive collection, PULSE records full-body optical motion capture with finger-level hand articulation, surface EMG, binocular eye tracking, wearable IMU, and a fingertip pressure array recording quantitative grip force, all sampled at 100 Hz. Each scenario is a long-horizon compositional task containing many smaller sub-tasks, yielding over 7,700 annotated action segments. Each segment is tagged with a motor primitive, the hand involved, the manipulated object, and a natural-language description with four paraphrased variants. On top of the dataset, we define multiple benchmark tasks for temporal cross-modal sensorimotor learning: scene and fine-grained action recognition, grasp onset anticipation, missing-modality robustness, tactile-driven sub-second grasp-state prediction, cross-modal pressure reconstruction, etc. Together these tasks evaluate how information transfers across physiology, motion, gaze, and touch, rather than reducing the dataset to an activity-label catalog. We evaluate three backbone architectures, nine fusion strategies, seven published baselines, and task-specific models including SyncFuse and DailyActFormer. Detailed task definitions and results are in the Appendices. Data, annotations, and baseline code will be publicly released.

Diffusion large language models (dLLMs) generate text by iteratively denoising masked token sequences. Although dLLMs can predict all masked positions in parallel within each step, the large number of denoising iterations still makes inference expensive. This cost can be reduced spatially by unmasking multiple tokens per step, or temporally by collapsing multiple denoising steps into one verification call. We propose Parallel Speculative Decoding (PSD), a training-free framework that jointly improves inference along both axes. Using the confidence scores from a single forward pass, PSD selects positions to unmask via a configurable, adaptive unmasking policy and constructs multi-depth speculative drafts without extra model calls. A final batched verification pass then applies hierarchical acceptance, keeping the deepest draft that remains consistent with the updated predictions. Experiments on three dLLMs across reasoning and code generation tasks show that PSD achieves favorable trade-offs between inference efficiency and generation quality, reaching up to $5.5\times$ tokens per forward pass with accuracy comparable to greedy decoding.