NeurIPS 2026 Papers

We propose **EverAnimate**, an efficient post-training method for long-horizon animated video generation that preserves visual quality and character identity. Long-form animation remains challenging because highly dynamic human motion must be synthesized against relatively static environments, making chunk-based generation prone to accumulated drift: (i) low-level quality drift, such as progressive degradation of static backgrounds, and (ii) high-level semantic drift, such as inconsistent character identity and view-dependent attributes. To address this issue, **EverAnimate** restores drifted flow trajectories by anchoring generation to a persistent latent context memory, consisting of two complementary mechanisms. *(i) Persistent Latent Propagation* maintains a context memory across chunks to propagate identity and motion in latent space while mitigating temporal forgetting. *(ii) Restorative Flow Matching* introduces an implicit restoration objective during sampling through velocity adjustment, improving within-chunk fidelity. With only lightweight LoRA tuning, **EverAnimate** outperforms state-of-the-art long-animation methods in both short- and long-horizon settings: at 10 seconds, it improves PSNR/SSIM by 8%/7% and reduces LPIPS/FID by 22%/11%; at 90 seconds, the gains increase to 15%/15% and 32%/27%, respectively. Anonymous project page: [https://anonymous-xxx-lab.github.io/](https://anonymous-xxx-lab.github.io/).

First-person dynamic spatial reasoning requires models to track continuous motion and precise geometric structure, but the quadratic attention cost of Transformer-based Video-LLMs makes dense visual tokens computationally expensive. Existing token pruning paradigms predominantly rely on discrete static snapshots, failing to preserve the physical dynamics essential for reasoning. We propose Event Cascade Pruning (ECP), to our knowledge the first training-free framework that leverages the high-frequency motion cues from event cameras as a continuous event-guided motion prior to guide token selection. ECP combines three stages: Event-Triggered Causal Sampling to anchor motion-informative keyframes, Event-guided Motion Saliency Filtering to remove static or low-motion regions, and Event-Attention Ranking Fusion to calibrate spatial attention with motion salient dynamics. With 80\% visual token reduction, ECP outperforms the full-token baseline (37.62\% vs. 36.31\%) while achieving 1.89× inference speedup and 52\% GFLOPs reduction. We further introduce ESR-Real, the first real-world RGB-event benchmark for first-person spatial reasoning, where ECP improves accuracy by 2.68\% over full-token baselines.

Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progression, or infer implausible relations across events. We study structured temporal understanding as the ability to recover event structure from visual evidence, including event boundaries, single-event progression, and multi-event relations. We argue that this gap arises from a mismatch between current post-training objectives and temporal reasoning, where caption-style supervision allows models to rely on language shortcuts rather than visual temporal grounding. To address this, we propose $\textbf{EventLens}$, a vision-centric reinforcement learning framework that learns video understanding through event-structure recovery}. Instead of reconstructing captions, EventLens trains models to recover temporal structure under controlled perturbations, such as altered temporal granularity, reversed progression, and shuffled event order. We instantiate this framework with a three-level temporal ontology and derive three verifiable task families: event segmentation, progression discrimination, and multi-event relation reasoning. These tasks admit deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models. We further introduce multi-teacher on-policy distillation (MT-OPD) to consolidate specialized policies into a unified model. Experiments on Qwen3-VL backbones show consistent improvements on temporal grounding and event-sensitive reasoning benchmarks, while preserving general video-QA performance. Learning curves further demonstrate that EventLens achieves comparable downstream transfer with significantly reduced training cost compared to direct mixed RL optimization.

Vision-Language-Action (VLA) policies translate language and visual inputs into robot actions, where their hidden representations directly shape closed-loop behavior. However, mechanistic interpretability tools from language and vision-language models do not transfer cleanly to VLAs: outputs are robot actions rather than human-readable tokens, and validating internal changes requires expensive closed-loop rollouts. We propose an event-grounded interpretability pipeline that anchors SAE feature analysis to behavioral events rather than text contexts. End-effector keyframes are clustered within each task using visual, state, and temporal cues, linking SAE features to behaviorally salient events and, via optional VLM annotations, to semantic context. To our knowledge, our pipeline is among the first to ground SAE-based VLA analysis in closed-loop behavioral events. Across two simulation architectures and a real-robot study, event-grounded ranking yields the strongest causal effects on OpenVLA and transfers to the continuous action chunks of $\pi_{0.5}$. SAE is a sparse but imperfect intervention basis: usability varies with architecture and intervention site, and aggressive intervention reveals safety and interpretability limits. Overall, event-grounded SAE analysis emerges as a practical starting point for behavior-anchored VLA interpretability, motivating future work on SAE features beyond action-aligned coordinates, finer-grained closed-loop evaluation, and safe interventions for high-stakes VLA deployments.

Real-world scenes often contain static, slow, and fast components within the same field of view. Frame cameras are inherently inefficient in this setting, as their global exposure and fixed sampling must follow the fastest motion and therefore oversample the rest of the scene. Event cameras provide asynchronous local sensing, but standard sensors use a single global contrast threshold, leading to a trade-off between sensitivity and redundancy: low thresholds preserve static information but over-trigger under fast motion, while high thresholds reduce data volume but lose slow or static structures. We propose a multi-threshold event imaging model that assigns motion-compatible thresholds to different scene components, enabling faithful multi-velocity-scale imaging with fewer events. We also introduce a practical acquisition pipeline that implements this model on existing event cameras. Experiments on synthetic and real scenes validate the effectiveness of our approach in preserving both static structures and fast dynamics while reducing data volume.

We introduce Sharper Transductive Local Complexity (STLC) as a new tool for analyzing the generalization performance of transductive learning methods, improving upon the current transductive bounds. Our work extends the classical local complexity-based analysis to the transductive setting, incorporating substantial and novel components beyond standard inductive and transductive analysis. Although Local Rademacher Complexity (LRC) has been used to obtain sharp inductive generalization bounds and local complexity-based transductive bounds, it has remained an open problem whether a localized Rademacher complexity framework can achieve exactly the same sharp bounds matching their inductive counterparts. STLC provides a confirmative answer to this question. STLC is constructed by first deriving a new and sharp concentration inequality for the supremum of empirical processes capturing the gap between test and training losses, or the test-train process, under uniform sampling without replacement. The proof establishes a Bernstein-type concentration inequality via a novel entropy-based approach built on the modified log-Sobolev inequality for the swap walk. A subsequent peeling strategy with a surrogate variance operator then yields excess risk bounds in the transductive setting that exactly match the classical LRC-based inductive bounds without the additional logarithmic gap in existing works. We further advance the current state-of-the-art in transductive learning through two applications: (1) for realizable transductive learning over binary-valued function classes with finite VC dimension $\dVC$ and $u \ge m \ge \dVC$, where $u$ and $m$ are the number of test features and training features, STLC gives a nearly optimal bound $\Theta(\dVC \log(me/\dVC)/m)$ nearly matching the minimax rate $\Theta(\dVC/m)$ up to $\log m$ and exactly matching the inductive bound, resolving a decade-old open question; and (2) STLC presents a sharper excess risk bound for transductive kernel learning compared to the prior local complexity–based results.

Vision-Language Models (VLMs) such as CLIP have demonstrated strong performance across tasks yet remain highly vulnerable to adversarial attacks. While adversarial fine-tuning has been demonstrated to be effective against such malicious inputs, its multi-step adversary generation scheme during fine-tuning further induces prohibitive computational costs for VLM backbones. Previous works in unimodal models have introduced single-step strategies to reduce this burden. However, we find naive single-step strategies fail on VLMs due to more severe catastrophic overfitting. In addition, we identify a novel failure mode termed *Perturbation Radius Overfitting*, where VLMs overfit to the specific training attack budget while paradoxically becoming fragile to weaker perturbations. We trace these failures to growing angular misalignment and magnitude surge of the local gradients, which compromise the local linearity essential for accurate single-step approximation. Guided by this analysis, we introduce a joint adversarial optimization scheme that actively rectifies the local geometry by suppressing both angular misalignment and magnitude surge along the perturbation path. Our approach consistently outperforms existing single-step methods across various datasets and VLM architectures, achieving state-of-the-art robustness while preserving efficiency. To our knowledge, this is the first work to systematically examine single-step adversarial fine-tuning on VLMs and establish the geometric foundations of their robustness.

When approximating an intractable density via variational inference *VI* the variational family is typically chosen as a simple parametric family that very likely does not contain the target. This raises the question: *Under which conditions can we recover characteristics of the target despite misspecification?* In this work, we extend previous theoretical results on robust VI with location-scale families under target symmetries in two substantial ways: (1) We open them up to a wider range of divergences by providing sufficient conditions for exact recovery of the target mean and correlation matrix when using the forward Kullback-Leibler divergence and $\alpha$-divergences. (2) By doing so, we find that we can drop the restrictive assumption of a log-concave target made in previous work, allowing us to give guarantees for a wider range of targets, including multi-modal ones. In our experiments, we show how our guarantees can serve as guidelines for the choice of the variational family and $\alpha$-value and we illustrate on a diverse set of examples how and why optimization can fail in the absence of our sufficient conditions.

This position paper argues that the evaluation of modern visual processing systems should no longer be driven primarily by single-metric image quality assessment benchmarks, particularly in the era of generative and perception-oriented methods. Image restoration exemplifies this divergence: while objective IQA metrics enable reproducible, scalable evaluation, they have increasingly drifted apart from human perception and user preferences. We contend that this mismatch risks constraining innovation and misguiding research progress across visual processing tasks. Rather than rejecting metrics altogether, this paper calls for a rebalancing of evaluation paradigms, advocating a more human-centered, context-aware, and fine-grained approach to assessing the visual models' outcomes.

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects (ω≤0.06), removing it causes small shifts (ω≤0.12) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points (ω≤0.31). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

Evaluation practices in machine learning are increasingly objects of scientific study in their own right: their assumptions, statistical properties, and failure modes determine which research conclusions can be trusted. We audit one such practice - the use of aggregate-subset Membership Inference Attacks (MIAs) for evaluating machine unlearning - across 8 dataset-architecture pairs, 4 unlearning methods, and over 500 experimental runs, identifying five failure modes: (1) limited and regime-dependent identifiability between baseline and retrained models, with KS-test indistinguishability ($p > 0.05$) in $\approx$42\% of experiment--task pairs; (2) regime-dependent signal-to-noise collapse; (3) monotonicity violations in 5 of 8 experiments under graded forgetting; (4) inter-task rank inconsistency, and (5) regime confounding by model training quality. To quantify these failures, we develop the Membership Inference Attack Unlearning Score (MIAU), a normalized diagnostic framework integrating three MIA comparisons as gap closure fractions between baseline and retrained references. We formalize four validity properties any unlearning metric should satisfy. Our findings indicate that aggregate-subset MIAs are reliable unlearning diagnostics only in narrow high-memorization regimes, and motivate complementary evaluation paradigms, alongside MIA-based protocols.

Access to real-world information is often noisy and fragmented. Constructing a coherent narrative from such fragments requires models to reconstruct missing spans within a broader storyline, commonly referred to as text infilling, while preserving consistency with both the local context and the global storyline. Despite using text infilling as a pre-training objective in many Large Language Models (LLMs), their actual performance on real-world narrative infilling remains underexplored. In this paper, we address this gap by introducing a multi-domain benchmark of ∼9.2K instances for narrative infilling, constructed by masking one to three sentences across four narrative types: encyclopedic text, commonsense stories, news articles, and visual narratives. Using this benchmark, we evaluate 20 instruction-tuned open-source LLMs ranging from 1.5B to 70B parameters across varying levels of instruction specificity and reasoning guidance. Outputs are assessed using standard automatic metrics and a qualitative framework covering five narrative dimensions. Results show that model scale does not reliably predict infilling quality: Gemma-2-2B achieves the highest qualitative score (4.02/5), outperforming models over ten times larger, including DeepSeek-Qwen-32B (3.77/5, ↓6.6%) and LLaMA-3.3-70B (3.71/5, ↓8.3%). We further find that explicit reasoning offers limited benefits as chain-of-thought reasoning yields only a marginal improvement (+0.6%). Additionally, short narratives and domain characteristics emerge as stronger predictors of task difficulty than infill position alone for narrative infilling in current LLMs.

Time series foundation models (TSFMs) are increasingly used for probabilistic forecasting across diverse domains, yet the reliability of their uncertainty estimates remains poorly understood. In this work, we study uncertainty calibration in probabilistic TSFMs from a broader methodological and empirical perspective. We introduce a taxonomy of how TSFMs represent predictive uncertainty and develop a fine-grained evaluation framework for calibration in multi-horizon forecasting that distinguishes pooled, horizon-wise, and stronger conditional notions of calibration. Using this framework, we conduct a comprehensive empirical evaluation of popular TSFMs across diverse datasets and show that current models exhibit distinct and systematic patterns of miscalibration that standard metrics often obscure. We further analyze common post-hoc recalibration methods and characterize which forms of calibration they can improve and which tend to persist. Our results provide practical guidance for evaluating and improving uncertainty calibration in TSFMs and highlight important limitations of current practice.

LLM agents are increasingly expected to operate as general-purpose systems that resolve real-world user requests, yet their dynamic scaling behavior in realistic environments remains poorly understood. In this paper, we systematically investigate two principal test-time scaling axes of LLM agents: sequential scaling through extended interaction and parallel scaling through trajectory sampling. We first introduce a realistic benchmark that provides one unified framework for evaluating LLM agents across search, coding, reasoning, and tool-use domains, more faithfully reflecting the heterogeneity of real-world deployments. Evaluating ten leading LLM agents reveals substantial performance degradation when transitioning from domain-specific evaluations to this realistic setting. Building on this foundation, we progressively scale test-time compute along fine-grained increments to characterize the performance upper bound. We find that neither scaling axis can consistently yield meaningful gains from additional test-time compute in realistic environments, a phenomenon we attribute to two fundamental limitations: the scaling plateau that bottlenecks sequential scaling and the verification gap that undermines parallel scaling. Code and data will be released publicly.

ECG, with mature physiological simulators and public benchmarks, is a useful testbed for evaluating what synthetic pretraining supports. We introduce a reusable framework for ECG representation learning that compares patient-free physiological simulators, patient-derived real ECGs, and learned synthetic ECGs from real-ECG-trained generators. Holding the encoder and MAE objective fixed, we study three matched settings. Transfer-protocol analysis compares abnormal-inclusive real ECGs, normal-only real ECGs, and simulator pretraining under frozen probing and full fine-tuning on 26 abnormal-ECG tasks from PTB-XL, G12EC, and CPSC2018 using a single-lead-II pipeline. Compute-normalized PTB-XL scaling tests whether simulator pretraining matches real-data pretraining under equal processed-sample budgets. A fixed 100k-sample generator-class comparison evaluates VAE, DCGAN, SSSD-ECG diffusion variants with and without labels, and two simulators, SimECG-M and SimECG-N. The transfer analysis shows that protocol determines the supported claim: abnormal-inclusive real ECG pretraining yields more linearly separable pathology features under frozen probing, whereas full fine-tuning narrows the gap and can reverse it when the simulator corpus is sufficiently scaled. Under the fixed-100k protocol, task-wise F1 averaged over five runs ranks SimECG-N first on 22/26 tasks; no evaluated learned generator ranks first under this fixed-100k protocol; and SimECG-N obtains higher mean F1 than both diffusion variants on all 26 tasks. These results delineate where abnormal-inclusive real exposure remains beneficial and where patient-free physiological simulation can serve as an effective, protocol-bound substitute under the evaluated single-lead-II MAE pretraining and full-fine-tuning setting. We release pretraining corpora, splits, and code to support reproducible evaluation of synthetic ECG pretraining sources.

Vision-Language Models (VLMs) have achieved strong performance on static visual understanding benchmarks, yet their ability to act in dynamic environments remains poorly understood. We introduce AtariBench, a controlled benchmark that evaluates VLMs across 20 Atari 2600 games, requiring models to make sequential decisions from high-temporal-resolution gameplay streams. By decoupling fast in-game dynamics from wall-clock inference speed, AtariBench focuses on whether models can perceive rapid visual state changes, reason over spatiotemporal dynamics, and ground actions in the current observation. Our experiments reveal substantial limitations in state-of-the-art VLMs: models often fail to infer object motion and direction, exhibit brittle sequential decision-making, and increasingly over-rely on textual interaction history rather than newly appended visual feedback. We further introduce a separability analysis to identify which in-game scores provide reliable signals of model competence, and use the benchmark to study how memory design and prompting strategies affect performance. Overall, AtariBench exposes a persistent gap between passive visual perception and active decision-making, providing a lightweight and rigorous testbed for developing VLMs that can reason and act in dynamic environments.

Physical reasoning is central to agents that step into the physical world. Models now excel at code and math, but their physical reasoning capabilities remain largely untested. Existing physical benchmarks evaluate goal-state matching rather than functional performance. This position paper argues that evaluating physical reasoning in LLM agents requires construction benchmarks, where physics verifies functional outcomes. To operationalize physical reasoning evaluation, we identify five capabilities spanning the full agentic loop from goal to verified artifact. Our audit of eleven physical benchmarks finds only two that combine realistic physics with functional verification, the minimum requirements for testing physical reasoning; both are construction benchmarks. Neither covers all five capabilities, yet frontier models already achieve near-zero success as task complexity increases. Progress depends on the breadth of benchmarks we build. We call on academia, game studios, and industry labs to expand construction benchmarks along both axes, widening capability coverage and deepening physical challenges. Their breadth determines how fully agents move from code and math to reshaping the physical world.

Foundation models of brain activity aim to learn representations from large-scale neural recordings that generalize across sessions, subjects, species, and downstream tasks. Transformer-based models have emerged as a promising architecture for this goal and begin all with the same first step --- tokenization: transforming sparse binary events into vector representations that downstream networks can consume. In other areas of machine learning, tokenization is its own design problem, distinct from the downstream foundation model. Several mostly implicit tokenization approaches dominate the literature, but their impact on downstream performance has not been systematically evaluated. We propose four explicit desiderata for a neural tokenizer---reconstruction fidelity, downstream utility, cross-instance generalization, and compression---and a framework of metrics spanning seven dimensions, including firing statistics, encoding and decoding, and latent structure and robustness. We assemble a benchmark of ten datasets spanning species, brain areas, recording technologies, and behavioral paradigms, and evaluate a representative set of discrete and continuous tokenizers, both single-neuron and population variants. Across the strategies evaluated, we show how tokenizers trained only on reconstruction can support downstream tasks. Yet reconstruction alone is misleading: reconstruction quality and downstream utility can be sharply dissociated. Our multi-axis framework is a response to this failure mode. These findings highlight tokenization as a critical and underexplored design axis in brain foundation models, and suggest that further progress will require continued systematic study of tokenization strategies.

Multimodal language models increasingly show promise for enabling the large-scale computational analysis of film, opening up new avenues for learning about film history and the evolution of narrative techniques. But the creation of stable benchmarks built around Hollywood films is complicated by copyright protections. In this work, we address these concerns directly, by building a new collection of Hollywood films defined by two criteria: box office popularity (where we publish the first large-scale, open collection of weekly box office earnings reported by Variety magazine from 1922-1979); and likely public domain status (by researching copyright registrations and renewals in the US Catalog of Copyright Entries). We build a new multimodal MCQ benchmark on top of this collection that focuses on narrative elements that directly evaluate the abilities of models to inform meaningful research on film narrative; we find that many vision-language models struggle on this task (with many performing at near-chance levels of accuracy), while audio-visual models (including those that use audio in captioning scenes) reach a maximum accuracy of 61.1%, well below human-level performance.