NeurIPS 2026 Papers

The rapid advancement of controllable video editing and agent-driven video generation has exposed a critical bottleneck: the inability of existing Multimodal Large Language Models to accurately perceive and articulate fine-grained differences between videos. Whether diagnosing editing fidelity or detecting training data quality issues, explicit comparative perception serves as an essential infrastructure. To address this capability gap, we introduce the ViDiC task and its corresponding ViDiC-1K benchmark. ViDiC-1K comprises 1,000 curated video pairs annotated with 3,720 fine-grained comparative checklist items, rigorously structured across seven dimensions: subject, style, background, camera work, motion, position, and playback techniques. To ensure reliable evaluation and prevent model hallucination, we propose a novel dual-checklist framework that assesses the accuracy of both similarities and differences separately based on an LLM-as-a-Judge protocol. Extensive experiments on 16 representative MLLMs reveal significant limitations in their comparative reasoning and dynamic difference perception. We hope ViDiC-1K serves as a foundational benchmark to advance continuous multi-video captioning and edit awareness in multimodal intelligence.

Training VideoLLMs for complex reasoning remains challenging due to sparse sequence-level rewards and the lack of fine-grained credit assignment over long, temporally grounded reasoning trajectories. While reinforcement learning with verifiable rewards (RLVR) provides reliable supervision, it fails to capture token-level contributions, leading to inefficient learning. Conversely, existing self-distillation methods offer dense supervision but lack structure and diagnostic specificity, and often interact unstably with reinforcement learning. In this work, we propose VISD, a structured self-distillation framework that introduces diagnostically meaningful privileged information for video reasoning. VISD employs a video-aware judge model to decompose reasoning quality into multiple dimensions, including answer correctness, logical consistency, and spatio-temporal grounding, and uses this structured feedback to guide a teacher policy for token-level supervision. To stably integrate dense supervision with RL, we introduce a direction–magnitude decoupling mechanism, where environment rewards determine update direction, while structured privileged signals modulate token-level update magnitudes. This design enables semantically aligned and fine-grained credit assignment, improving both reasoning faithfulness and training efficiency. Additionally, VISD incorporates independent advantage estimation, curriculum scheduling, and EMA-based teacher stabilization to support robust optimization over long video sequences. Experiments on various benchmarks demonstrate that VISD consistently outperforms strong baselines across diverse video reasoning tasks, achieving gains in accuracy, grounding quality, and interpretability. Notably, VISD significantly accelerates convergence, highlighting the effectiveness of structured self supervision in improving both performance and sample efficiency for VideoLLMs.

Recent advances in video generation have made prompt-based control increasingly central to AIGC video generation. Prompts specify what a video should depict and how it should be represented, controlling factors such as visual style or camera behavior. Understanding this recoverability is important both for creative reuse and editing, and for assessing prompt leakage risks. However, existing video understanding benchmarks do not measure this capability: a caption may describe what is visible, but a replayable prompt must recover the generation-relevant controls needed to reproduce the video. To address this gap, we introduce VI-Bench, a benchmark built from 16.1 million real-user prompts and 900 human-verified AIGC videos. VI-Bench spans three progressively harder settings, namely single-shot semantic grounding, control over style and camera behavior, and multi-shot compositional inversion, and evaluates five generation-critical dimensions: subject, action, scene, style, and camera. We evaluate 18 representative VLMs, including 2 proprietary and 16 open-source models on VI-Bench, using an Inversion Score that measures prompt-level alignment with the original prompt and video-level fidelity of the regenerated video. The results reveal substantial limitations: even the strongest model achieves only 0.632 on Inversion Score, performance degrades sharply as samples require richer control and multi-shot reasoning, and models often produce plausible prompts whose regenerated videos deviate from the reference. These findings show that video prompt inversion is a distinct and under-evaluated capability requiring models to transform visual understanding into replay-stable generative control.

Large-scale video–text contrastive learning has become the de facto standard for video understanding, yet mainstream methods almost universally adopt the linear patch embedding of vision Transformers (ViT) or evenly sliced tubelets as the front end. We revisit this front end from the biological principles of the primary visual cortex (area V1) and introduce V1DynamicVisionSystem (V1-DVS), a fully end-to-end trainable, bio-plausible video embedding framework. Its main contributions are: (i) a three-pathway parallel architecture that maps onto layer 4C orientation selectivity, layer 4B motion direction selectivity, and the cytochrome-oxidase Blob color-opponent system; (ii) an explicit shape–color decoupling, with isotropic Blob processing running in parallel with anisotropic Gabor orientation processing; (iii) long-range spatiotemporal convolution that captures inter-frame optical flow through 3D filters and the Adelson–Bergen motion-energy model rather than naive frame stacking; and (iv) bidirectional cross-pathway attention, allowing motion and color features to mutually modulate, as in V1→V4 inter-areal feedback. Pretrained on CC3M, WebVid-10M, ActivityNet and Ego4D and evaluated under a fully self-supervised objective(SimCLR + temporal-order verification + masked-token prediction), V1-DVS outperforms four strong baselines—linear patch embedding, mean temporal pooling, tubelet embedding, and spatiotemporal latent encoding—on six metrics spanning retrieval and embedding-space geometry, validating the practical value of biologically grounded structure for video embedding.

Understanding 4D scenes is a core challenge in visual intelligence, requiring models to jointly capture scene geometry, appearance, and dynamics from partial observations. Existing approaches, however, are subdivided into feedforward models that are used for direct geometric prediction and generative models for visual synthesis. We propose a unified generative-predictive framework for 4D scene understanding that jointly models RGB video and geometric representations such as depth and camera pose with a single joint generative model. Our generative formulation enables us to learn from diverse, heterogeneous datasets with different subsets of geometric and visual annotations, and enables us to condition on arbitrary subsets of inputs, allowing a single model to perform cross-modal generation, geometric prediction, and future scene inference from partial observations. In addition, our generative approach enables test-time search over multiple candidate scene completions, improving geometric prediction from sparse observations. Overall, we illustrate the efficacy of our generative-predictive framework on a suite of generation and prediction tasks.

We introduce $\textbf{UniCon3R}$, a unified feed-forward framework for online human-scene 4D reconstruction from monocular video. Current feed-forward human-scene reconstruction methods suffer from artifacts, where bodies float above the ground or penetrate parts of the scene. A key reason is the lack of effective interaction modelling between the human and the environment. Our goal is to exploit contact between the human and the scene during inference to actively improve the human mesh reconstruction. To that end, we explicitly model interaction by inferring 4D contact from the human pose and scene geometry and use the contact as a corrective cue for generating the pose. This enables UniCon3R to jointly recover scene geometry and spatially aligned 4D humans within the scene. Experiments on standard human-centric video benchmarks show that UniCon3R outperforms state-of-the-art baselines on physical plausibility and global human motion estimation while preserving fast, feed-forward inference speeds. The results validate our central claim: contact serves as a powerful internal prior, thus establishing a new paradigm for physically grounded joint human-scene reconstruction. Source code and models will be released upon acceptance.

Unified Multimodal Models (UMMs) excel in general tasks but struggle to bridge the gap between personalized understanding and generation. Prior works largely rely on implicit token-level alignment via supervised fine-tuning, which fails to fully capture the potential synergy between comprehension and creation. In this work, we propose Sync-R1, an end-to-end reinforcement learning framework that jointly optimizes personalized understanding and generation within a single, explicit reasoning loop. Through this unified feedback process, Sync-R1 enables personalized comprehension to guide content creation, while the resulting generation quality reciprocally refines understanding within an integrated reward landscape. To efficiently orchestrate this dual-task synergy, we introduce Sync-GRPO, a reinforcement learning method utilizing an ensemble reward system. Furthermore, we propose Dynamic Group Scaling (DGS), which adaptively filters low-potential trajectories to reduce gradient variance and accelerate convergence. To better reflect real-world complexity, we introduce UnifyBench++, featuring denser textual descriptions and richer user contexts. Experimental results demonstrate that Sync-R1 achieves state-of-the-art performance, showcasing superior cross-task reasoning and robust personalization without requiring complex cold-start procedures.

Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data. We show that CoT generalizes by repeatedly applying a shared single-step predictor to compose reusable ID transitions. Theoretically, the OOD task risk is controlled by the sum of ID subtask risks, with an error-propagation coefficient independent of reasoning depth/steps $T$. A Rademacher-complexity analysis further gives a generalization gap of $\mathcal{O}(\sqrt{T/n})$, rather than linear in $T$, due to predictor sharing, where $n$ denotes training samples per subtask. For autoregressive Next-Token Prediction, we derive an explicit bound $\mathcal{O}\big(k\big(\sqrt{{T}/{mn}}+\sqrt{{T}/{n}}+{T}/{m}\big)\big)$, where $m$ and $k$ denote context length and output token number per step. Experiments on Atom-Task and GSM8K support the theory, showing that CoT enables short-to-long and compositional generalization, while explicit state tracking is crucial.

We investigate foundational questions about the interplay between memorization and learning in large language models (LLMs). When an LLM generates a string, can we disentangle the roles of rote memorization and contextual learning? Rote memorization results in an LLM regurgitating specific details of a training string, while contextual learning leads to the generation of new strings that follow generalizable patterns of the underlying language. Related questions of interest include: can an LLM avoid memorization when optimally learning a language?, and if not, can we design training schemes to reduce memorization and improve learning? To address these questions, we propose contextual memorization, a new measure that precisely disentangles memorization from learning. Using this measure, we establish that different training strings are memorized at different epochs in training, and that memorization of some training strings is unavoidable when optimally learning a language. Finally, we propose memorization-aware training, a scheme that dynamically down-samples training strings based on their memorization, and in doing so, equalizes learning across strings, reduces memorization, and improves learning. Our conclusions are supported by extensive experiments on multiple LLMs using formal languages as a controlled testbed, enabling precise language specification, exact string sampling, and elimination of data contamination; and further validated on natural languages.

A central challenge in understanding deep sequence models is characterizing how token representations evolve across layers: whether they collapse, diverge, or converge to nontrivial structures. We study this question for Transformers with hyper-connections, whose learnable cross-layer coupling makes the stability of representation collapse analytically tractable. In contrast to standard residual connections, hyper-connections introduce additional parameters that directly control the hyperbolicity of the zero fixed point, at which the local stability is fully determined by the Jacobian eigenvalues, thereby enabling a rigorous and precise bifurcation analysis of the dynamical system underlying the model. In particular, we derive a continuous-depth ODE limit of self-attention with hyper-connections and use bifurcation theory to characterize its token dynamics. For the 1-stream case, we obtain a closed-form critical threshold at which a pitchfork bifurcation occurs, separating representation collapse from divergence or convergence to nontrivial stable states. For the 2-stream case, the richer coupling structure yields two distinct stability boundaries, corresponding to qualitatively different instabilities: one arising when a real eigenvalue crosses zero, altering the number of nearby equilibria (static type), and another when a complex-conjugate pair reaches the imaginary axis, leading to oscillatory behavior (Hopf type). These phenomena have no counterpart in the 1-stream setting. Experiments on pretrained hyper-connected language models validate our theory and show that unstable regimes yield stronger performance.

Algorithm unrolling is ubiquitous in machine learning, particularly in hyperparameter optimization and meta-learning, where Jacobians of solution mappings are computed by differentiating through iterative algorithms. Although unrolling is known to yield asymptotically correct Jacobians under suitable conditions, recent work has shown that the derivative iterates may initially diverge from the true Jacobian, a phenomenon known as the curse of unrolling. In this work, we provide a non-asymptotic analysis that explains the origin of this behavior and identifies the algorithmic factors that govern it. We show that truncating early iterations of the derivative computation mitigates the curse while simultaneously reducing memory requirements. Finally, we validate our theoretical findings both on synthetic problems and in a meta-learning setting for few-shot classification, where we study the behavior of gradients obtained via unrolling and implicit differentiation.

Spiking Neural Networks (SNNs) offer high energy efficiency, but direct optimization is challenging because spike generation is governed by a non-differentiable hard threshold. Practical direct-training methods, including surrogate gradients (SG) and local zeroth-order (LocalZO) estimators, maintain hard spikes in the forward pass while using surrogate or perturbation-induced derivatives in the backward pass. This forward--backward mismatch raises a central theoretical question: how do local gradient discrepancies propagate through the spatiotemporal computation graph, and can the resulting biased updates still provide guaranties for the original discrete-spike objective? We develop a unified convergence analysis framework for SG and LocalZO direct SNN training. A smoothed spike objective is introduced only as an analytical bridge, allowing us to separate the core error into two components: the gradient-level mismatch between the direct-training mean direction and the smoothed-objective gradient, and the objective-level consistency gap between the smoothed and original discrete objectives. Both components are controlled by a threshold-tail functional that measures how often hard-trajectory membrane potentials lie near the firing threshold. Under suitable regularity assumptions, we obtain a non-asymptotic guarantee for the original discrete-spike objective. Experiments across representative SNN architectures and benchmarks verify convergence consistency, the existence of errors, and demonstrate that the threshold-tail statistic is an effective diagnostic for direct SNN training.

Besides domination-based algorithms such as the NSGA-II, the decomposition-based MOEA/D is one of the most popular and successful optimization heuristic for multi-objective problems. It decomposes the problem into several, single-objective subproblems. By solving these in a co-evolutionary manner and archiving non-dominated solutions, it approximates or computes the Pareto front. Interestingly, so far no advantages from the co-evolutionary communication of subproblem solutions were proven. In this first mathematical runtime analysis addressing this research question, we investigate how the mutation-based MOEA/D solves the classic OMM benchmark when solutions are exchanged among subproblems in different distances $r$. We show, for the first time, that the MOEA/D profits from an exchange of solutions, and that already very small communication radii $r$ are sufficient to maximally profit from this feature of the algorithm. More precisely, for all values $r \ge 1$ we prove a tight runtime guarantee of $\Theta(n^2)$, where $n$ is the problem size. This is asymptotically faster than the known bound of $O(n^2 \log n)$ guarantee for $r=0$, which we complement with a matching bound of $\Omega(n^2 \log n)$.

Generative optimization uses large language models (LLMs) to iteratively improve artifacts (like code, workflows or prompts) using execution feedback. It is a promising approach to building self-improving agents, yet in practice remains brittle: despite active research, only 9% of surveyed agents used any automated optimization. We argue that this brittleness arises because, to set up a learning loop, an engineer must make "hidden" design choices: What can the optimizer edit and what learning evidence is provided per update? We investigate three factors that affect most applications: the starting artifact, batching execution traces as experiences, and determining a credit horizon with truncated traces. Through case studies in MLAgentBench, Atari, and BigBench Extra Hard, we find that decisions around these factors can determine whether generative optimization succeeds and they are not made explicit in prior works. Different starting artifacts determine which solutions are reachable in MLAgentBench, truncated traces can still improve Atari agents, and larger minibatches do not monotonically improve generalization on BBEH. We conclude that the lack of a simple, universal learning-loop setup across domains is a major hurdle for productionization and adoption, and provide practical guidance for making these design choices.

This position paper argues that understanding generalization in diffusion models requires fundamentally new theoretical frameworks that go beyond both classical statistical learning theory and the benign overfitting paradigm developed for supervised learning. In diffusion models, unlike in supervised learning, memorization of training data and generalization to novel samples are incompatible: a model that has fully memorized its training set generates copies rather than novel data. Several theoretical explanations for why practical diffusion models nevertheless generalize have been proposed, based on capacity limitations, implicit regularization from optimization, or architectural inductive biases, but their interactions remain unclear. We argue that the field should pivot from explaining why the diffusion models do not memorize to investigating what the models actually learn during pre-memorization phase. To highligh our stance, we conduct empirical study of diffusion models trained on CIFAR-10, and we distill the findings into concrete open questions that we believe are key to improve understanding of generalization in diffusion models.

Recent advances in Time Series Foundation Models (TSFMs) increasingly rely on axial attention mechanisms, but despite their empirical success, the theoretical properties of axial attention remain largely unexplored. To bridge this gap, we extend the Single Location Regression framework introduced in Marion et al. [2025] to 2D structured data, in order to study sparse information settings where a target depends on a small number of tokens. We formulate an analytically tractable axial attention-type predictor and rigorously analyze its statistical properties and training dynamics. We prove that under a specific asymptotic regime and with knowledge of the underlying oracle parameters, our predictor achieves Bayes optimality, whereas a class of semi-linear simplifications strictly fails to do so, underlining the role of the inner non-linearity. We prove that for a specific temperature scaling, Projected Gradient Descent converges globally to the optimal oracle parameters. Extensive numerical experiments validate our theoretical findings and highlight the critical role of inverse temperature scheduling in empirical convergence.

Graph Neural Networks (GNNs) have demonstrated remarkable effectiveness on graph-based tasks. However, their predictive confidence is often miscalibrated, typically exhibiting under-confidence, which harms the reliability of their decisions. Existing calibration methods for GNNs normally introduce additional calibration components, which fail to capture the intrinsic relationship between the model and the prediction confidence, resulting in limited theoretical guarantees and increased computational overhead. To address this issue, we propose a simple yet efficient graph calibration method. We establish a unified theoretical framework revealing that model confidence is jointly governed by class-centroid-level and node-level calibration at the final layer. Based on this insight, we theoretically show that reducing the weight decay of the final-layer parameters alleviates GNN under-confidence by acting on the class-centroid level, while node-level calibration acts as a finer-grained complement to class-centroid level calibration, which encourages each test node to be closer to its predicted class centroid at the final-layer representations. Extensive experiments validate the superiority of our method.

Large vision-language models (VLMs) often exhibit degraded safety alignment when visual inputs are integrated. Even when a text prompt is explicitly harmful, adding an image can substantially increase the jailbreak success rate. In this paper, we observe that VLMs can clearly distinguish benign, refusal, and jailbreak samples in their representation space, and that jailbreak responses often contain safety warnings. These observations lead to the \textit{recognize-but-fail-to-refuse} hypothesis: VLM jailbreaks do not arise from a failure to recognize harmful intent, but instead occur because adding an image induces a representation shift that steers the sample into a distinct jailbreak state where refusal is not triggered. To quantify how this image-induced representation shift contributes to jailbreak behavior, we define a jailbreak direction and characterize the jailbreak-related representation shift as the projection of the total image-induced shift onto this direction. Our analysis shows that the jailbreak-related shift reliably characterizes jailbreak behavior, providing a unified explanation for diverse jailbreak scenarios. Finally, we propose JRS-Rem, a defense method that enhances VLM safety by removing the jailbreak-related shift at inference time. Experiments show that JRS-Rem significantly improves VLM safety across multiple scenarios while preserving utility on benign tasks.

Predictive Coding (PC) is an influential account of cortical learning. Much of recent work has focused on comparing PC to Backpropagation (BP) to find whether PC offers any advantages. Small scale experiments show that PC enables learning that is more sample efficient and effective in many contexts, though a thorough theoretical understanding of the phenomena remains elusive. To address this, we quantify the efficiency of learning in BP and PC through a metric called "target alignment", which measures how closely the change in the output of the network is aligned to the output prediction error. We then derive and empirically validate analytical expressions for target alignment in Deep Linear Networks. We show that learning in PC is more efficient than BP, which is especially pronounced in deep, narrow and pre-trained networks. We also derive exact conditions for guaranteed optimal target alignment in PC and validate our findings through experiments. We study full training trajectories of linear and non-linear models, and find the predicted benefits of PC persist in practice even when some assumptions are violated. Overall, this work provides a mechanistic understanding of the higher learning efficiency observed for PC over BP in previous works, and can guide how PC should be parametrised to learn most effectively.