NeurIPS 2026 Papers

Burst imaging under unstable illumination suffers from flicker, a periodic degradation that modulates local brightness differently across adjacent frames, creating frame-specific flicker states. Existing methods rely on spatial priors or reference-centric fusion without explicitly modeling the evolving flicker state, whose temporal periodicity makes it predictable from adjacent observations by world models that capture latent state transitions. Motivated by this predictive capacity, we propose StateFlicker, a state-aware burst restoration framework that interprets adjacent frames as observations under evolving flicker states. StateFlicker comprises two state-aware designs: Cross-State Visibility (CSV) and Flicker-State Understanding (FSU). CSV leverages complementary visibility across adjacent flicker states to selectively recover content suppressed under the center state. With this cross-state support, FSU constructs dual state pathways through a frozen world model to obtain predicted and observed center states, and distills their state discrepancy into dense features that characterize the center-specific flicker condition for restoration. Extensive experiments on the burst flicker removal dataset show that StateFlicker achieves superior performance over state-of-the-art methods.

Equipping artificial agents with spatial intelligence requires a comprehensive generative prior over the dynamic 3D world. We propose World Motion Models (WMMs) that capture "what was, is, and will be where across time" via sparse SE(3) pose trajectories. WMMs are built on the observation that elements of dynamic scenes can be well-approximated by a set of rigid SE(3) trajectories, a minimal yet expressive primitive for 4D modeling. This representation unifies articulated objects, human bodies, hand-object interactions, piecewise-rigid scene dynamics, camera motion, and even robot states and actions into a single shared space. Given this representation, we cast the joint distribution of these entities as a flexible sequence modeling problem, utilizing flow-matching with per-token noise levels. Coupled with a context token mechanism with non-sequential conditioning, this formulation supports any-to-any marginal conditioning across an arbitrary number of entities and time steps. Tasks such as future prediction, motion infilling, model-predictive control, inverse kinematics, cross-embodiment retargeting, and policy learning all reduce to the application of different masks over the same network. Experiments on diverse 6 applications of 3D vision and robotics demonstrate the versatility and flexibility of WMMs with strong performance.

RLHF is widely used to align flow-matching text-to-image models with human preferences, but often leads to severe diversity collapse after fine-tuning. In RL, diversity is often assumed to correlate with policy entropy, motivating entropy regularization. However, we show this intuition breaks in flow models: policy entropy remains constant, even while perceptual diversity collapses. We explain this mismatch both theoretically and empirically: the constant entropy arises from the fixed, pre-defined noise schedule, while the diversity collapse is driven by the mode-seeking nature of policy gradients. As a result, policy entropy fails to prevent the model from converging to a narrow high-reward region in the perceptual space. To this end, we introduce perceptual entropy that captures diversity in a perceptual space and maintains the property of standard entropy. Building upon this insight, we propose two entropy-regularized strategies, Perceptual Entropy Constraint and Perceptual Constraints on Generation Space, to preserve perceptual diversity and improve the quality. Experiments across two base models, neural and rule-based rewards, and three perceptual spaces demonstrate consistent gains in the quality-diversity trade-off; PEC achieves the best overall score of 0.734 (vs.\ baseline's 0.366); a complementary setting of PEC further reaches a diversity average of 0.989 (vs.\ baseline's 0.047). Code will be released.

Compositional inference - the decomposition of observations into an unknown number of latent components - is central to perception and scientific data analysis. Attention-based models perform well when components are approximately separable, as in object-centric vision. Under additive superposition, however - where multiple components contribute to every observation - we identify a structural failure mode we term slot collapse: multiple slots converge to the same dominant component while weaker ones remain unrepresented. We trace this to a general limitation: attention is memoryless with respect to explained evidence. All slots repeatedly operate on the same input without accounting for what has already been explained, so gradients are dominated by the strongest component, inducing shared fixed points across slots. As a result, attention fails to enforce non-redundant allocation under additive superposition. We address this by introducing residual evidence modeling, which tracks the remaining explanatory capacity of the input, instantiated via evidence depletion - a minimal modification combining multiplicative depletion with an attention bias. Controlled ablations show that parallel attention, sequential processing alone, and loss-based regularization all fail to resolve collapse; evidence depletion, which adds stateful residual tracking to sequential attention, consistently succeeds. Across synthetic benchmarks and real-world audio mixtures (FUSS), evidence depletion reduces slot collapse by up to an order of magnitude, generalizing beyond synthetic settings. On gravitational-wave source inference for the ESA/NASA LISA mission, under identical architectures, data, and losses, standard attention fails while evidence depletion prevents collapse and enables multi-source posterior estimation. These results show that under additive superposition, residual evidence tracking is the operative ingredient - preventing collapse and enabling compositional inference.

Recent advances in Multi-modal Large Language Models (MLLMs) have significantly improved video understanding. However, real-world scenarios often require going beyond isolated video inputs, giving rise to the emerging paradigm of video deep research, which involves open-world reasoning through integration of external knowledge. Despite this progress, existing approaches are limited by their inability to perform effective inter-video interaction and their reliance on static, one-shot datasets that yield noisy and suboptimal supervision. To address these challenges, we propose VideoSailor, an open-world video deep research agent powered by a trajectory-to-policy flywheel that iteratively converts generated reasoning trajectories into high-quality off-policy guidance to mitigate the limitations of low-quality on-policy rollouts. For systematic evaluations, we further introduce Omni-BrowseComp, a benchmark that emphasizes omni-modal evidence aggregation, cross-video reasoning, and fine-grained temporal grounding with reasoning-relevant segment annotations. Extensive experiments demonstrate that VideoSailor significantly improves performance on complex multi-hop and cross-video reasoning tasks, establishing a scalable and effective paradigm for open-world video deep research.

Generative models, such as diffusion and flow-based models, have emerged as powerful policy approximators in deep reinforcement learning (RL) due to their ability to represent multi-modal action distributions. However, their potential for maximum entropy RL remains largely untapped, as most existing methods treat generative processes as black-box samplers or rely on heuristic noise. In this paper, we propose GENIEO, a novel off-policy generative RL framework that leverages the exact intrinsic entropy of an augmented dummy-action policy to drive principled exploration. By designing a structurally invertible affine dual-variable flow, we bypass the numerical approximations of standard ODE solvers. This innovation allows us to rigorously apply the change-of-variables formula to derive a tractable and exact formulation for the augmented dummy-action policy entropy, enabling its direct integration into the off-policy learning objective. Experimental evaluations across 14 challenging tasks from the DMControl and HumanoidBench suites demonstrate that GENIEO achieves state-of-the-art (SOTA) performance. By effectively discovering optimal strategies in complex, high-dimensional environments, our approach bridges the gap between expressive generative modeling and maximum entropy RL.

Genomic events span multiple spatial scales, from single-nucleotide substitutions to large structural variants, yet existing foundation models process only reference DNA sequences, leaving methylation, short variants, and structural-variant boundaries as downstream labels or auxiliary inputs. We propose the Unified Genomic Model (UGM), an encoder-only genomic foundation model that learns multi-scale genomic events as native tokens within a unified 43-token vocabulary. UGM introduces three algorithmic components: (i) a unified event tokenizer that maps reference bases, SNPs, CpG methylation, short INDELs, and structural-variant boundaries into a single event-state space; (ii) event-balanced pretraining, which promotes adequate gradient coverage for rare but biologically important events; and (iii) Site-Level Joint Prediction (SJP), a masked-language-modeling objective that recovers the complete genomic state at a masked position as a single token rather than predicting base, methylation, and variant labels independently. We evaluate UGM against specialized and generalist baselines across functional variant classification, regulatory benchmarks (NT and GUE), structural-variant detection, and cross-event attention analysis. UGM achieves competitive performance, particularly on tasks where explicit event-state information is central. It obtains the highest score on all seven functional variant tasks, reaches competitive splicing prediction on NT, and notably outperforms DNA-only baselines on SV breakpoint detection. These results suggest that native multi-scale event pretraining is a promising direction for genomic representation learning.

Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust Region Policy Distillation (TOP-D), which transforms the notoriously unstable, high-variance On-Policy Distillation (OPD) into a stable training paradigm by dynamically constructing a proximal teacher. Theoretically, we establish a rigorous framework demonstrating that TOP-D inherently controls gradient variance. By providing a formal global convergence analysis alongside a monotonic improvement bound, we mathematically formalize the reliability and stability of the overall training dynamics. Empirically, TOP-D dramatically enhances training stability, sample efficiency, and final performance on mathematical reasoning tasks. More importantly, TOP-D introduces zero additional computational overhead, positioning itself as a promising alternative to the well-established OPD paradigm.

Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average, surpassing Gemini-2.5 Pro in standard agent-workflow setting (37.9%). At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image-bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.

Financial markets are complex, noisy environments which present unique challenges for representation learning and world modeling. In this study, we systematically explore the application of supervised and self-supervised representation learning methods to financial markets data and their ability to learn world models of financial markets. To support this, we assemble and release $\textbf{Market-1T}$ a dataset of more than one trillion observations spanning all US equities from 2008 through 2025. We then develop a domain-specific data augmentation and apply a variety of modern SSL objectives including Joint Embedding Predictive Architecture (JEPA), Masked Autoencoder (MAE), and DINO. Our results highlight that supervised methods are still dominant in this domain when applied directly to core quantitative finance tasks: predicting changes in prices, volatility, and transactions cost. Further, when supervised models are trained across tasks they are able to leverage complementarities which enhance cross-task performance. While performance of SSL-based methods on these core tasks lags behind, we find that they better capture latent structure---such as time and assets effects---which supervised models miss. Our results suggest current SSL objectives capture broad features of markets but further work is needed to close the gap between supervised methods.

Diffusion language models (DLMs) promise parallel, order-agnostic generation, but on standard benchmarks they have historically lagged behind autoregressive models in sample quality and diversity. Recent continuous flow-matching approaches over one-hot and embedding spaces have narrowed this gap, suggesting continuous state spaces are highly effective for language. In this work, we further close the autoregressive gap by modeling text as a continuous diffusion process over fixed-width bitstreams. Our approach represents semantic tokens as analog bit sequences and utilizes a matched-filter residual parameterization to isolate contextual learning from analytic independent-bit posteriors. Crucially, our empirical comparisons reveal that while deterministic continuous DLMs are competitive, they can be overly contractive, undershooting real-data entropy. We demonstrate that the remaining quality-diversity gap is bridged by a stochastic sampler that applies Langevin-type corrections gated by the entropy-rate profile, automatically concentrating stochasticity in high-information regions while remaining nearly deterministic elsewhere. On the One Billion Word Benchmark (LM1B), our 130M-parameter bitstream model reaches a generative perplexity (GenPPL) of 59.76 at matched real-data entropy (4.31) using 256 neural function evaluations (NFEs), decisively outperforming prior DLM baselines and reaching the autoregressive reference. On OpenWebText (OWT), our stochastic sampler establishes a new continuous-DLM Pareto frontier, achieving GenPPL = 27.06 at an entropy of 5.26 using 4x fewer steps than previous 1024-NFE baselines. As an additional architectural benefit, bitstream diffusion removes the O(V) vocabulary scaling bottleneck shared by standard DLMs. By predicting O(log V) bitwise logits via semantic bit-patching, our model yields a reduced memory footprint and higher throughput, demonstrating a scalable paradigm for language generation as vocabulary sizes grow.

Multivariate time series forecasting is important in real-world applications, but practical data often show both temporal and spectral non-stationarity. Existing methods may treat severe spectral drift as informative evolution, allow unreliable spectral components to contaminate temporal representations through credibility-agnostic time--frequency interaction, and infer cross-variable dependencies from noisy local correlations, which may lead to spurious long-term dependencies. To address these issues, we propose TFNS, a dual-stream time--frequency framework for suppressing unreliable spectral evidence under non-stationarity. Specifically, we introduce a Spectral Credibility Estimator to quantify patch-level spectral stability and softly suppress low-credibility frequency components for more reliable frequency modeling; a Credibility-Gated Cross-Interaction module to enable bidirectional time--frequency enhancement while adaptively constraining cross-domain information injection according to spectral reliability; and a Spectral Cointegration Graph to capture stable long-term cross-variable structure in the spectral domain while mitigating spurious dependencies under non-stationary drift. Together, these modules form a progressive pipeline from local spectral purification to reliability-aware cross-domain interaction and robust long-term structure modeling. Experiments on 11 benchmarks show that TFNS achieves state-of-the-art performance, ranking first in 44/55 MSE and 43/55 MAE comparisons.

Recent multimodal large language models (MLLMs) increasingly rely on visual chain-of-thought to perform region-grounded reasoning over images. However, existing approaches ground regions via either textified coordinates—causing modality mismatch and semantic fragmentation—or fixed-granularity patches that both limit precise region selection and often require non-trivial architectural changes. In this paper, we propose Numerical Visual Chain-of-Thought (NV-CoT), a framework that enables MLLMs to reason over images using continuous numerical coordinates. NV-CoT expands the MLLM action space from discrete vocabulary tokens to a continuous Euclidean space, allowing models to directly generate bounding-box coordinates as actions with only minimal architectural modification. The framework supports both supervised fine-tuning and reinforcement learning. In particular, we replace categorical token policies with a Gaussian (or Laplace) policy over coordinates and introduce stochasticity via reparameterized sampling, making NV-CoT fully compatible with GRPO-style policy optimization. Extensive experiments on three benchmarks against eight representative visual reasoning baselines demonstrate that NV-CoT significantly improves localization precision and final answer accuracy, while also accelerating training convergence, validating the effectiveness of continuous-action visual reasoning in MLLMs. The code is provided in the Supplementary Materials.

On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensitive to teacher choice and loss formulation, whereas OPSD fails in our tested settings due to test-time absence of instance-specific privileged information (PI). In contrast, OPSD is effective when PI represents a shared latent rule, such as a system prompt or alignment preference. We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.

Bayesian experimental design (BED) has traditionally been based on maximising expected uncertainty reductions from prior to posterior. A major shortfall of this approach is that it leads to doubly intractable objectives that are difficult to optimise, while customising them to particular downstream tasks of interest can also be difficult. Following first principles decision theory, we demonstrate that BED can alternatively be formulated in terms of an expected future loss (EFL) on downstream actions, providing a simple and naturally task-driven framework. Critically, we then show that all such EFLs can be rearranged into *singly intractable objectives* that can be jointly optimised with respect to both the design policy and a downstream *action policy* using stochastic gradient descent methods. This formulation further sidesteps the need for any explicit posterior or marginal likelihood estimation and is naturally implicit, requiring only the ability to sample from the joint model over model parameters and data, and evaluate the downstream loss function. It thus allows design policies to be learned more effectively, efficiently, and simply than existing methods, while providing easy customisation to different downstream tasks and losses.

We increasingly use machine learning to label scientific datasets. The models we develop and deploy are improving all the time, but they are not and will likely never be perfect. Mistakes matter, as errors can propagate into our scientific understanding, particularly when systematically biased. Very reasonably, scientists thus review substantial proportions of ML-generated labels to verify or correct mistakes in pursuit of ensuring their scientific findings are not biased by ML. In this work, we focus on helping scientists optimally allocate this reviewing effort relative to their scientific goals. We focus on a specific class of scientists (ecologists) and a specific, widespread, and impactful modeling target (occupancy modeling, which estimates where species are likely to occur, conditioned on environmental factors). We introduce ACORN (Active Continuous-Score Occupancy Modeling), a method that incorporates ML predictions into occupancy models and strategically selects samples for expert review that are maximally informative for downstream ecological analysis. Across camera-trap and bioacoustic datasets, our method recovers ecological conclusions close to those obtained from fully human-labeled data, while requiring substantially fewer expert reviews than non-targeted review policies. Our results suggest that ML-assisted scientific workflows should optimize expert effort for downstream inference, rather than for classifier accuracy alone, especially when human review budget is limited.

Verifiable tool-interactive agents produce rich trajectories that contain far more information than final success or failure. However, standard RL post-training with verifiable rewards often collapses these trajectories into scalar outcomes and trains over a largely fixed data stream, leaving open the question of which examples are most useful for improving the current policy. We introduce TReDS, a trajectory-grounded framework for capability-conditioned training distribution shaping. TReDS estimates multidimensional task requirements from offline probing trajectories and controlled tool-view evidence, maintains the policy capability state in the same semantic space, and uses reliability-gated online correction to form a policy-conditioned requirement representation. A scheduler then converts requirement--capability mismatch, recent utility, and stabilizing pressure into a time-varying training distribution for GRPO, without modifying the GRPO objective. We evaluate TReDS with a progressive stress-test suite spanning $\tau^2$-Bench official evaluation, $\tau$-Bench native-bridge evaluation, BFCL V3, and API Bank, using only the $\tau^2$-Bench training split for policy optimization. TReDS improves over the GRPO baseline under the same training data and backend, and shows stronger behavior retention under same-family and cross-protocol evaluations. These results suggest that RL post-training for tool-interactive agents should optimize not only how trajectories are rewarded, but also where trajectory-revealed requirements allocate training probability mass.

Backtesting large language models on historical events requires reasoning exclusively from information available before a specified cutoff date. Yet models routinely leak post-cutoff knowledge from pre-training into their reasoning, inflating apparent accuracy and undermining evaluation validity. Prompt-based constraints fail when suppressed content is causally related to the prediction, and knowledge unlearning cannot address this problem because temporal compliance is instance-specific: the same fact may be legitimate evidence for one cutoff date and a violation for another. Rather than erasing knowledge, the model must learn **temporal discipline**: selecting evidence conditioned on each instance's cutoff date. We propose **TEMPO** (**T**emporal **E**nforcement via **M**ode-separated **P**olicy **O**ptimization), which trains this discipline via two contributions: (1) a two-mode reward where a leakage mode drives post-cutoff claims to zero as a hard prerequisite before a performance mode optimizes task performance; and (2) a GRPO-based training pipeline that enables the model to discover temporally valid reasoning strategies. We prove that training monotonically decreases leakage, converges to the leak-free optimum, and improves task performance once compliance is achieved. On three prediction tasks and two models, TEMPO reduces leakage from 2$\sim$13\% to 0.6$\sim$3.7\% across all conditions, with task performance improving 6$\sim$13\% where strong pre-cutoff signals exist and maintained where the prediction task is inherently difficult from valid information alone.

Scaling robot policy learning for autonomous surgery is challenging, as expert demonstrations are expensive and in vivo exploration poses substantial safety risks. Surgical world models address this by generating realistic, action-conditioned future frames from an initial observation, but existing methods exhibit two persistent failure modes: *spatial interaction incoherence*, where visible instrument contact fails to induce spatially consistent tissue deformation, and *temporal fidelity collapse*, where prediction errors compound across autoregressive rollouts and progressively corrupt visual quality. We present **SurgVista**, a surgical world model that mitigates both failures through two training recipes. *Deformation Consistency Regularization* extracts scene-point trajectories from training videos and enforces cross-frame coherence through latent contrastive learning, strengthening physically consistent instrument-tissue dynamics. *Drift Adaptation Training* mitigates long-horizon drift by perturbing conditioning frames with online prediction residuals and photometric augmentations calibrated to long-horizon drift statistics, sustaining visual fidelity over extended rollouts. To enable rigorous evaluation, we further introduce **SurgWorld-Bench**, featuring diverse procedure types, long-range rollouts, and decoupled metrics for instrument-motion accuracy and tissue-response fidelity. Extensive experiments show that SurgVista consistently outperforms state-of-the-art methods across visual quality, temporal consistency, and interaction fidelity, with gains widening as the prediction horizon grows.