NeurIPS 2026 Papers

Inverse Reinforcement Learning (IRL) aims to recover a reward function that explains expert demonstrations. Existing IRL methods typically rely on a bi-level optimization procedure that alternates between reward learning and policy optimization, leading to substantial computational burden and training instability. In this work, we introduce a different route that eliminates policy optimization entirely by leveraging diffusion policies. Our key insight is that a diffusion policy encodes the action-gradient structure of the optimal soft Q function, enabling reward learning to be cast as a sequence of value recovery problems, thereby allowing us to bypass reward-policy loops inherent in prior IRL methods. Specifically, our method proceeds in three stages: (I) recovering the optimal soft Q function via action-gradient matching and estimating the corresponding soft value function (LogSumExp of Q values) in a way inspired by Gumbel regression; (II) calibrating these soft values by inferring a state-dependent offset; (III) extracting the reward by enforcing Bellman consistency. This leads to *Loop-Free Inverse Reinforcement Learning* (LFIRL), a fully offline algorithm that operates in a simple, loop-free, and sequential manner. LFIRL is simple to implement, significantly improves training efficiency while maintaining strong reward recovery performance. Empirically, across Maze, Franka Kitchen, Adroit Hand Pen, and Push-T benchmarks, LFIRL achieves **2-3x** speedup over the fastest baselines, while matching or surpassing state-of-the-art methods in reward recovery quality.

This position paper argues that evaluation in generative molecular dynamics (GenMD) exhibits a recurring pattern of comparison against insufficient or misleading benchmarks, which risks over-claiming advantage over classical baselines. The pattern reflects a structural misalignment between common evaluation practices and the scientific tasks the field sets for itself. The reason is structural: the natural ground truth, exhaustive molecular dynamics on the systems of interest, is computationally costly, the very gap that motivates the field. With ground truth inaccessible, evaluation drifts toward convenient proxies, and the resulting norms entrench easily. We identify misaligned evaluation patterns that persist as a community-level equilibrium rather than failures of individual papers, and propose concrete observables, reporting standards, and benchmark designs, both computational and experimental, that better reflect what GenMD aspires to deliver. Our aim is constructive: to give the field shared language and tools for evaluation that match its scientific ambitions before today's conventions calcify into norms.

We revisit regularized regret minimization under full-information and bandit feedback, where a learner optimizes an objective of the form $\langle r, \pi \rangle - \eta^{-1} \psi(\pi)$ for adversarially chosen rewards $r$, policies $\pi \in \Delta_K$, and a known strongly-convex regularizer $\psi(\cdot)$. We establish the minimax regularized regret for a broad class of regularizers, specifically those that are relatively strongly-convex and smooth with exponent $p \geq 1$, Legendre, and coordinate-wise separable. For large $\eta$, the problem approaches unregularized online learning, recovering the standard $\Theta(\sqrt{T})$ scaling. For small $\eta$, the strong convexity of the regularizer dominates, and we identify the regimes that attain $\Theta(\eta \log T)$ minimax regret under both feedback models. Notably, while prior logarithmic regret lower bounds are largely restricted to specific curved loss functions (e.g., squared loss), our lower bounds hold for ***any*** such regularizer against purely linear adversarial rewards. A key technical contribution is our proof strategy, which leverages a Bregman geometric perspective of regularized regret over the quotient space $\mathbb{R}^K / \mathrm{span}(\\{\mathbf{1}\\})$ to resolve the translation invariance of the inverse mirror map. By combining this dual geometric formulation with the ***van Trees inequality***, we reduce the adversarial online learning to Bayesian estimation, offering a unified framework for establishing logarithmic regret lower bounds.

3D object understanding and generation methods in indoor scenes produce impressive results, yet they often overlook a pervasive source of information in real-world scenes: repeated objects. We introduce the task of lookalike object detection in indoor scenes, which leverages repeated and complementary cues from identical and near-identical object pairs. Given an input scene, the task is to classify pairs of objects as identical, similar or different using multiview images as input. To address this, we present Lookalike3D, a multiview image transformer that effectively distinguishes such object pairs through similarity learning by harnessing strong semantic priors from large image foundation models. To support this task, we collected the 3DTwins dataset, containing 76k manually annotated identical, similar and different pairs of objects based on ScanNet++, and show an improvement of 14\% IoU over baselines on the task of lookalike object detection. We also demonstrate how our method improves the downstream tasks of joint 3D object reconstruction and part co-segmentation, turning repeated and lookalike objects into a powerful cue for consistent, high-quality 3D scene understanding. Our code, dataset and models will be made publicly available.

Transformers dominate video recognition. They split videos into tokens, and processing them has expensive superlinear computational cost. Yet videos are filled with redundancy, so we can question the need for this expense. We introduce LookWhen, a selector–extractor framework that factorizes video recognition into learning when, where, and what to compute. Our shallow selector gets a scaled-down video and quickly scores all tokens across space-time, while our deep extractor gets the top-K selected tokens to approximate full-video representations without actually processing all the tokens. A key challenge is defining effective supervision for selection and extraction. For selection pre-training, we introduce a score on representations that ranks tokens by uniqueness using a simple nearest-neighbor distance. For extraction pre-training, we distill both a video teacher and an image teacher, for which we normalize its frame-wise representations to learn what changes within videos. Through these strategies, our selector-extractor learns general and efficient representations for feature extraction or fine-tuning to a task. Through experiments on Kinetics-400, SSv2, Epic-Kitchens, Diving48, Jester, and Charades, we show that LookWhen achieves a better accuracy-computation trade-off than efficient models and upgraded baselines of similar size. LookWhen Pareto-dominates in accuracy-FLOPs on 9 of 12 cases (6 tasks X 2 settings) and roughly matches on 3. In accuracy-throughput, measuring time in practice, LookWhen is more efficient still at 6.7$\times$ faster than InternVideo2-B at equal accuracy.

Essential visual information can be sparse, yet most visual computation remains dense. Though only a fraction matters, vision transformers process images as uniform sets of tokens, scaling cost with their number. As these models grow stronger, they grow larger, and the cost of computing everything climbs higher. Adaptive computation has shown that we can get by with less, but existing methods struggle at extreme sparsity and are difficult to _specialize_ to a task. We address this limitation by introducing _LookThere_, training an efficient selector of locations and an expressive extractor of representations to derive meaning from images _without ever processing the full high-resolution input_. We do so by learning the selector and extractor end-to-end with actor-critic reinforcement learning. The selector learns _where_ to look, the extractor _what_ to see, together saving computation by selecting only what is worth processing for a given task. We show that _LookThere_ selects the task-specific input, excelling at sparse recognition in high-resolution settings (traffic signs, billiards), and maintaining accuracy with as little as 0.2% of the input. It also generalizes across tasks and models, including global recognition (ImageNet classification), local recognition (ADE20K segmentation), zero-shot classification (by distillation), and regression (class agnostic enumeration). Across all settings, _LookThere_ surpasses state-of-the-art selection, providing a general and scalable framework for specialized and efficient adaptive computation.

Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions. Recent *thinking with images* approaches mitigate this limitation by allowing models to inspect images through operations such as zooming, cropping, or code execution. While effective, their multi-turn interactions with VLMs introduce redundant token computation and slow inference, and separate region inspection from the reasoning step that depends on it. To address these issues, we propose **Lo**ok Before You **Re**ason (LoRe), an implicit visual thinking framework for efficient visual reasoning within a single-turn VLM interaction. Given an image and a question, LoRe predicts the most relevant image region, aligns it with the corresponding image patch features, and strengthens the associated key-value (KV) cache entries, enabling subsequent decoding to directly exploit the selected visual cues without additional visual inputs or explicit tool calls. We further employ reinforcement learning to guide the VLM toward identifying informative image regions and leveraging them during reasoning. To support fine-grained evaluation of multimodal reasoning, we introduce COCO-ObjQA, a large-scale benchmark for object-centric multimodal reasoning. Experiments on three challenging multimodal benchmarks demonstrate that LoRe consistently improves VLM reasoning performance with low additional inference overhead.

Despite the remarkable progress of large Vision-Language Models (VLMs) in Radiology Report Generation (RRG), existing black-box autoregressive generation paradigms lack intermediate clinical logical reasoning. This deficiency renders such models prone to generating linguistically fluent reports with erroneous diagnostic findings, which manifest as plausible hallucinations and thereby undermine clinical reliability. To this end, we propose a novel RRG paradigm termed "Look-Before-You-Leap", aimed at enhancing the clinical logical reasoning capabilities of the model to improve diagnostic accuracy. Specifically, we first construct a self-evolving Radiological Skill (RadSkill) repository to guide the teacher model in distilling high-confidence explicit clinical reasoning, and further internalize this reasoning capability into the implicit diagnostic intuition of the RRG baseline. Building upon this foundation, a novel preference optimization method termed Psychometric Preference Optimization (PSY-PO) is proposed. Grounded in Item Response Theory (IRT) from psychometrics, this method precisely samples clinically misleading dispreferred reports for preference optimization, thereby effectively mitigating plausible hallucinations. Extensive experimental results demonstrate that our proposed method significantly enhances both the generation quality and trustworthiness of radiology reports.

Video-based world models hold significant potential for generating high-quality embodied manipulation data. However, current video generation methods struggle to achieve stable long-horizon generation: classical diffusion-based approaches often suffer from temporal inconsistency and visual drift over multiple rollouts, while autoregressive methods tend to compromise on visual detail. To solve this, we introduce LongScape, a hybrid framework that adaptively combines intra-chunk diffusion denoising with inter-chunk autoregressive causal generation. Our core innovation is an action-guided, variable-length chunking mechanism that partitions video based on the semantic context of robotic actions. This ensures each chunk represents a complete, coherent action, enabling the model to flexibly generate diverse dynamics. We further introduce a Context-aware Mixture-of-Experts (CMoE) framework that adaptively activates specialized experts for each chunk during generation, guaranteeing high visual quality and seamless chunk transitions. Extensive experimental results demonstrate that our method achieves stable and consistent long-horizon generation over extended rollouts. Our code is available at: https://anonymous.4open.science/r/AMSVVD-fdg245.

Long video out-painting is an important yet underexplored problem, while existing video out-painting methods are mainly developed for relatively short clips. However, extending video out-painting to long videos is highly challenging. A natural solution is iterative out-painting, \ie, generating videos chunk by chunk, but this introduces two major issues: unstable even crashed long-horizon generation caused by inter-chunk inconsistency and accumulated errors, and semantic drift over long temporal horizons. In this paper, we propose LongSCOP, short for Semantically Consistent long Video Out-Painting. First, we provide a fundamental training recipe with three components for smooth and stable long-horizon generation: self-forcing training, latent calibration, and future references. Built upon this stable generation foundation, a more advanced requirement is semantic consistency, for which we further develop a VLM-based agent that automatically selects the most informative reference frames from the whole video, enabling coherent subjects, scenes, and backgrounds throughout long-range generation. To support training and evaluation, we construct a large-scale, high-quality training dataset and introduce LVO-Bench, a dedicated benchmark curated by human experts for assessing long-horizon stability and semantic consistency. Extensive experiments demonstrate clear advantages of our method over existing approaches in both generation stability and semantic coherence, significantly improving the quality and reliability of long video out-painting.

Agents in real-world settings operate over long and evolving horizons, where information is repeatedly updated and can interfere with each other across memories, requiring accurate recall and aggregated reasoning over multiple pieces of information. However, existing benchmarks focus on static, independent recall and fail to capture these dynamic interactions between evolving memories. In this paper, we study how current systems perform in realistic, continuously evolving, long-horizon settings across diverse domains and question types. To this end, we construct an analytical benchmark, LongMINT (Long-Horizon Memory under INTerference), which features (1) long, highly interconnected contexts with frequently updated information, (2) diverse coverage across multiple memory domains (Wikipedia, code, multi-turn dialogue, and state tracking), enabling evaluation of domain generalization, and (3) diverse question types, including (i) single-target recall tasks that test retrieval under interference over long contexts and (ii) multi-target aggregation tasks that require counting, ordering, or reasoning across multiple relevant pieces of information. We evaluate over six representative systems, including vanilla long-context LLMs, retrieval-augmented generation methods, and memory-augmented agent frameworks. We observe consistently low performance (avg. 27.7\% accuracy), especially on questions that require aggregated reasoning over multiple pieces of evidence. Fine-grained analysis shows that performance is primarily limited by retrieval and memory construction capabilities. Furthermore, current memory systems struggle to recall and reason over facts that are multiple steps back, and performance decreases when this lookback distance increases. These findings highlight the need for more robust memory management systems for dynamic, long-horizon environments across varying domains.

Embodied Visual Tracking (EVT) aims to autonomously control an agent to follow a target in 3D space, which is critical for robotic navigation and human-robot interaction. However, robust long-term EVT in real-world scenarios faces two key bottlenecks: frequent tracking failures caused by occlusions and distractors, and the lack of a long-term benchmark for model training and evaluation. To address these, we propose a systematic solution. First, we develop LT-VLA, a lightweight VLA model featuring a Target Memory module to maintain consistent target identification and an Adaptive Execution module to adjust tracking actions based on observation reliability. Second, we construct LTEVT, a unified long-term EVT benchmark comprising 11 high-fidelity indoor and outdoor environments, enriched with over 5,000 assets and 325 human targets to simulate occlusions and distractors. Along with the benchmark, we release the LTEVT-4500K dataset for large-scale training and a comprehensive evaluation toolkit. Extensive experiments demonstrate that our 0.6B model achieves state-of-the-art performance comparable to 7B-scale models while maintaining superior efficiency. Notably, LT-VLA achieves 62.4% average Success Rate on EVT-Bench and runs at 20 FPS on a Unitree G1 robot, delivering a practical and robust solution for real-world deployment.

Algorithmic risk predictors are increasingly being used to inform resource allocation in healthcare and other high-stakes domains. Common in many such settings, the public decision maker delegates service provision to private organizations and compensates them according to each individual's predicted cost of service. We show that this common allocation mechanism can fail when organizations engage in \emph{favorable selection}, i.e., strategically choosing individuals whose actual costs are below the payment model predictions. Motivated by health insurance payments in the Medicare Advantage (MA) program, we provide a theoretical framework to study the interaction between a decision maker and a strategically selecting service provider. Our theoretical framework predicts two empirical patterns observed in Medicare Advantage that have become central concerns: enrollment in the private program expands, while government payments rise above the counterfactual cost of direct public provision. Guided by our theoretical framework, we then propose a reformed payment policy that is implementable under the government's data-access constraints. We prove that its equilibria eliminate overpayment while preserving profits for efficient organizations. We validate the framework in synthetic and semi-synthetic experiments that illustrate the advantages of our reformed policy, over the status quo.

While AI weather models excel at short-to-medium range forecasts (up to 15 days), they frequently suffer from ill-defined "instabilities" when rolled out over longer horizons. This work addresses the lack of a formal taxonomy by categorizing these failures into three distinct regimes — blow-up, drift, and loss of seasonality — through year-long rollouts of nine state-of-the-art AI weather models. Our analysis reveals that stability hinges on the treatment of small spatio-temporal scales: unstable models amplify high-frequency energy, while stable models act as denoisers when noise is added to their inputs. Far from reducing these models to mere stochastic parrots, our findings highlight that stable models generate unique weather trajectories, conditioned on the initial state. We verify our findings through ablation studies on architectural design choices, conducted using state-of-the-art Vision Transformer (ViT) AI weather model architectures.

Graph Neural Networks (GNNs) are powerful tools for learning from spatio-temporal data, as interactions in space can be naturally described by a graph structure. However, capturing long-range dependencies becomes substantially harder when information must flow through space and time simultaneously. Most existing approaches extend GNNs from static graphs to temporal settings by alternating propagation between time and space. In this work, we introduce STORM, a differential-equation-inspired GNN that exploits oscillatory dynamics to propagate information effectively in the joint spatio-temporal domain. By combining a wave-equation update with dissipative and external forcing terms, STORM balances conservative and non-conservative dynamics. We provide a bottom-up analysis of the model, highlighting its propagation behavior and stability, and show that STORM is universal. We empirically validate our method on diverse benchmarks, including tasks designed for analyzing long-range spatio-temporal dependencies, real-world forecasting, and a new long-range task inspired by physical simulations. Across these settings, STORM consistently matches or outperforms strong baselines, establishing a new state of the art for long-range spatio-temporal graph learning.

Off-policy, value-based reinforcement learning methods such as Q-learning are appealing because they can learn from arbitrary experience, including data collected by older policies or other agents. In practice, however, bootstrapping makes long-horizon learning brittle: estimation errors at later states propagate backward through temporal-difference (TD) updates and can compound over time. We propose *long-horizon Q-learning* (LQL), which introduces a principled backstop against compounding error when learning the optimal action-value function. LQL builds on a prior *optimality tightening* observation: any realized action sequence lower-bounds what the optimal policy can achieve in expectation, so acting optimally earlier should not be worse than following the observed actions for several steps before switching to optimal behavior. Our contribution is to turn this inequality into a practical stabilization mechanism for Q-learning by using a hinge loss to penalize violations of these bounds. Importantly, LQL computes these penalties using network outputs already produced for the TD error, requiring no auxiliary networks and no additional forward passes relative to Q-learning. When combined with multiple state-of-the-art methods on a range of online and offline-to-online benchmarks, LQL consistently outperforms both 1-step TD and n-step TD learning at similar runtime.

Long-horizon agency requires an AI system to preserve task state across hours or days of work, many tool calls, and repeated context resets. A common response is to scale a single resource: the model’s context window, together with inference-time reasoning. We argue that this response conflates long horizon with long context. Durable long-horizon state should not live only in the prompt; it should be externalized into the harness, the non-model infrastructure that manages tools, files, plans, checkpoints, sub-agents, permissions, and recovery. Publicly documented long-running agent systems already rely on such structures, including progress files, plan files, git history, rollback checkpoints, scratchpads, sub-agent decomposition, and browser-state checkpoints. We support this position with three lines of argument. First, memory-hierarchy history shows that pressure to scale one working resource repeatedly gave way to externally managed hierarchical state. Second, we identify five structural mismatches between long-context scaling and long-horizon agency: degraded effective context, lossy compaction, a single trust region, no transactional semantics, and no working-set decomposition. Third, we map long-horizon requirements to what context scaling provides and what harness-managed state must supply. We conclude by proposing eight harness state-management patterns and arguing that harness state should become a first-class research object for long-horizon agency.

Sparsity has long been a central theme in LLM efficiency but its role in context processing remains unresolved. As LLM workloads shift toward longer contexts and agentic interactions, the compute and memory bottlenecks of attention become increasingly critical, raising the question of whether these constraints are fundamental. Our position: this constraint is artificial, unnecessary and the future of LLM inference lies in extreme but principled sparsity along the context dimension. Our position comes out of various bits of empirical and theoretical evidence. Firstly, we find insistance on dense attention unreasonable since in long context a query essentially projects the O(N) attention information into hidden space of diemsion $d << N$. This, as we show is extremely lossy. Instead, we propose a completely context-sparse context processing combining top-k sparsity with linear attention / SSMs as the way forward. To support our proposal we show that this combination can approximately cover the functional space of dense softmax attention. Independently but importantly, in a first elaborate study of its kind, we empirically show a strong trend towards current LLM models, which are not trained for context sparsity, being extremely robust to inference time decode-sparsity across tasks of varying complexities such as retrieval, multi-hop QA, and mathematical reasoning and agentic coding. For instance, Qwen3.5-27B can tolerate upto 100x sparsity on benchmarks of RULER-HARD, LOFT and AIME2025 without loss of quality, and upto $50\times$ on SWE with a small drop in quality. These results emphasize the possibility that we can transition to complete sparsity without any loss of capability. Additionally, we also discuss what this shift in paradigm of context processing means for hardware. Importantly, we show that even current hardware is equipped enough to realise gains from this sparsity. For instance, our sparse decode kernels can accelerate large context processing by a factor 10x over FlashInfer at 50x sparsity levels on current hardware such as H100. Overall, these results position extreme context sparsity not as a heuristic, but as a principled foundation for LLM inference, training, and architecture design, both feasible and beneficial, and a compelling direction for future systems.

Long-form generation in language models often collapses into repetitive loops well before the nominal context limit. This is commonly attributed to model capacity or training data, and addressed through costly retraining or post-training pipelines. We show that the sampling algorithm is a primary determinant of long-context output quality, improving outputs without further training costs. Across ten open-weight models from 1B to 120B parameters, output lengths from 8K to 64K tokens, and more than 13,500 generations, full-distribution-aware samplers (P-less, Top-H, top-$n\sigma$) dominate automated metrics and human evaluation, while standard truncation (top-$p$, top-$k$) collapses into late-context repetition; min-$p$ falls in a middle tier. A 1,500-annotation Amazon Mechanical Turk study finds P-less winning more than nine in ten head-to-head pairings against a top-$p$ 0.9 baseline on a 120B model, with 10-gram repetition approaching zero at 64K tokens. A per-step selected-token rank diagnostic isolates the mechanism: standard truncation drives the rank to the argmax floor by the final quartile, while full-distribution-aware methods stay in a healthy band across the full window. This gap that widens with model scale, establishing sampler choice as an under-appreciated bottleneck for long-context generation in open-weight models. We frame long-context generation as sampling from outside the training distribution, and encourage further study for usecases like long-context agentic coding.