ICLR 2026 Papers

Temporal Video Grounding (TVG) aims to localize video segments corresponding to a given textual query, which often describes human actions. However, we observe that current methods, usually optimizing for high temporal Intersection-over-Union (IoU), frequently struggle to accurately recognize or understand the underlying actions in both the video and query, thus reducing the effectiveness of these methods. To address this, we propose a novel TVG framework that integrates inversion-based TVG as auxiliary objectives to maintain the model's action understanding ability. We introduce three kinds of inversion TVG tasks derived from the original TVG annotations: (1) Verb Completion, predicting masked verbs (actions) in queries given video segments; (2) Action Recognition, identifying query-described actions; and (3) Video Description, generating descriptions containing query-relevant actions given video segments. These inversion tasks are entirely derived from the original TVG tasks and are probabilistically integrated with them within a reinforcement learning framework. By leveraging carefully designed reward functions, the model preserves its ability to understand actions, thereby improving the accuracy of temporal grounding. Experiments show our method outperforms state-of-the-art approaches, achieving a 7.1\% improvement in R1@0.7 on Charades-STA for a 3B model.

Real-world videos often show routine activities punctuated by memorable, surprising events. However, most Video-LLMs process videos by sampling frames uniformly, likely missing critical moments that define a video's narrative. We introduce SPIKE, an inference-time framework that quantifies Bayesian Surprise as the belief update triggered by new visual evidence in the video stream, identifying moments where new visual evidence conflicts with prior beliefs. SPIKE effectively localizes surprise in videos, correlated with humans on positive (FunQA) and negative (Oops!) surprise benchmarks. SPIKE-RL further improves on SPIKE's ability to detect surprise, leveraging GRPO to refine its belief hypotheses based on a reward signal from the video caption. SPIKE and SPIKE-RL guide query-agnostic surprise-weighted frame sampling, which allocates more frames to interesting moments in the video. With this strategy, we achieve consistent performance gains on five downstream benchmarks. By enabling Video-LLMs to track beliefs and register surprise, our work paves the way for more robust models that can revise their understanding in response to new information.

Video Large Language Models (VideoLLMs) extend the capabilities of vision-language models to spatiotemporal inputs, enabling tasks such as video question answering (VideoQA). Despite recent advances in VideoLLMs, their internal mechanisms on where and how they extract and propagate video and textual information remain less explored. In this study, we investigate the internal information flow of VideoLLMs using mechanistic interpretability techniques. Our analysis reveals consistent patterns across diverse VideoQA tasks: (1) temporal reasoning in VideoLLMs initiates with active cross-frame interactions in early-to-middle layers, (2) followed by progressive video-language integration in middle layers. This is facilitated by alignment between video representations and linguistic embeddings containing temporal concepts. (3) Upon completion of this integration, the model is ready to generate correct answers in middle-to-late layers. (4) Based on our analysis, we show that VideoLLMs can retain their VideoQA performance by selecting these effective information pathways while suppressing substantial amount of attention edges, e.g., 58% in LLaVA-NeXT-7B-Video-FT. These findings provide a blueprint on how VideoLLMs perform temporal reasoning and offer practical insights for improving model interpretability and downstream generalization.

The growing scale of deep learning models has rendered exhaustive hyperparameter (HP) optimization prohibitively expensive. A promising solution is the use of scale-aware HPs, which can enable direct transfer of optimal settings from small-scale grid searches to large models with minimal performance loss. Such approaches are useful when the optimal settings converge "fast" enough with scale. While approaches like the Maximal Update Parameterization ($\mu$P) have empirically displayed fast transfer when scaling model width, a deeper conceptual understanding of the mechanisms that enable this is still missing. Our work establishes a systematic conceptual framework for analyzing fast HP transfer across different synthetic and practical scenarios. In synthetic settings, we present various quantitative examples where transfer either offers a provable computational advantage or fails even under $\mu$P. We then propose a key property that enables the fast transfer often observed in practice: through a novel decomposition of the optimization trajectory, we identify one component that rapidly converges with model width and determines the optimal HPs, and the other that continues to improve the loss with increased width but has negligible impact on HP choice. We conjecture that this decomposition elucidates the key mechanisms behind fast transfer and empirically validate it in practical settings such as LLM training.

Urban development impacts over half of the global population, making human-centered understanding of its structural and perceptual changes essential for smart city planning. While Multimodal Large Language Models (MLLMs) have shown remarkable capabilities across various domains, existing benchmarks that explore their performance in urban environments remain limited, lacking systematic exploration of temporal evolution and subjective perception of urban environment that aligns with human perception. To address these limitations, we propose UrbanFeel, a comprehensive benchmark designed to evaluate the performance of MLLMs in urban development understanding and subjective environmental perception. UrbanFeel comprises 14.3K carefully constructed visual questions spanning three cognitively progressive dimensions: Static Scene Perception, Temporal Change Perception, and Subjective Environmental Perception. We collect multi-temporal single-view and panoramic street-view images from 11 representative cities worldwide, and generate high-quality question-answer pairs through a hybrid pipeline of spatial clustering, rule-based generation, model-assisted prompting, and manual annotation. Through extensive evaluation of 20 state-of-the-art MLLMs, we observe that Gemini-2.5 Pro achieves the best overall performance, with its accuracy approaching human expert levels and narrowing the average gap to just 1.5%. Most models perform well on tasks grounded in scene understanding. In particular, some models even surpass human annotators in pixel-level change detection. However, performance drops notably in tasks requiring temporal reasoning over urban development. Additionally, in the subjective perception dimension, several models reach human-level or even higher consistency in evaluating dimension such as beautiful and safety. Our results suggest that MLLMs are demonstrating rudimentary emotion understanding capabilities. Our UrbanFeel benchmark will be made publicly available.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose **ParallelBench**, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We are releasing our benchmark to help accelerate the development of truly efficient dLLMs.

Controllable video generation remains a significant challenge, despite recent advances in generating high-quality and consistent videos. Most existing methods for controlling video generation treat the video as a whole, neglecting intricate fine-grained spatiotemporal relationships, which limits both control precision and efficiency. In this paper, we propose \textbf{Co}ntrollable \textbf{V}ide\textbf{o} \textbf{G}enerative \textbf{A}dversarial \textbf{N}etworks (\ourmes) to disentangle the video concepts, thus facilitating efficient and independent control over individual concepts. Specifically, following the \textbf{minimal change principle}, we first disentangle static and dynamic latent variables. We then leverage the \textbf{sufficient change property} to achieve component-wise identifiability of dynamic latent variables, enabling independent control over motion and identity. To establish the theoretical foundation, we provide a rigorous analysis demonstrating the identifiability of our approach. Building on these theoretical insights, we design a \textbf{Temporal Transition Module} to disentangle latent dynamics. To enforce the minimal change principle and sufficient change property, we minimize the dimensionality of latent dynamic variables and impose temporal conditional independence. To validate our approach, we integrate this module as a plug-in for GANs. Extensive qualitative and quantitative experiments on various video generation benchmarks demonstrate that our method significantly improves generation quality and controllability across diverse real-world scenarios.

Video understanding, including video captioning and retrieval, is still a great challenge for video-language models (VLMs). The existing video retrieval and caption benchmarks only include short descriptions, limits their ability of detailed video understanding evaluation. To address this problem, we present CaReBench, a testing benchmark for fine-grained video Captioning and Retrieval with 1,000 high-quality pairs of videos and human-annotated detailed captions. Uniquely, it provides manually separated spatial annotations and temporal annotations for each video. Based on this design, we introduce two evaluation metrics, ReBias and CapST, specifically tailored for video retrieval and video captioning tasks, respectively. These metrics enable a comprehensive investigation into the spatial and temporal biases inherent in VLMs. In addition, to handle both video retrieval and video captioning tasks in a unified framework, we develop a simple baseline based on a Multimodal Language Model (MLLM). By implementing a two-stage Supervised Fine-Tuning (SFT), we fully unlock the potential of MLLM, enabling it not only to generate detailed video descriptions but also to extract video features. Surprisingly, experimental results demonstrate that, compared to the CLIP-based models designed for retrieval and the popular MLLMs skilled in video captioning, our baseline shows competitive performance in both fine-grained video retrieval and video detailed captioning.

Text-to-video diffusion models have enabled high-quality video synthesis, yet often fail to generate temporally coherent and physically plausible motion. A key reason is the models' insufficient understanding of complex motions that natural videos often entail. Recent works tackle this problem by aligning diffusion model features with those from pretrained video encoders. However, these encoders mix video appearance and dynamics into entangled features, limiting the benefit of such alignment. In this paper, we propose a motion-centric alignment framework that learns a disentangled motion subspace from a pretrained video encoder. This subspace is optimized to predict ground-truth optical flow, ensuring it captures true motion dynamics. We then align the latent features of a text-to-video diffusion model to this new subspace, enabling the generative model to internalize motion knowledge and generate more plausible videos. Our method improves the physical commonsense in a state-of-the-art video diffusion model, while preserving adherence to textual prompts, as evidenced by empirical evaluations on VideoPhy, VideoPhy2, VBench, and VBench-2.0, along with a user study.

Mixture-of-Experts (MoE) has become a cornerstone in recent state-of-the-art large language models (LLMs). Traditionally, MoE relies on $\mathrm{Softmax}$ as the router score function to aggregate expert output, a designed choice that has persisted from the earliest MoE models to modern LLMs, and is now widely regarded as standard practice. However, the necessity of using $\mathrm{Softmax}$ to project router weights into a probability simplex remains an unchallenged assumption rather than a principled design choice. In this work, we first revisit the classical Nadaraya–Watson regression and observe that MoE shares the same mathematical formulation as Nadaraya–Watson regression. Furthermore, we show that both feed-forward neural network (FFN) and Mixture-of-Experts (MoE) can be interpreted as a special case of Nadaraya–Watson regression, where the kernel function corresponds to the input neurons of the output layer. Motivated by these insights, we propose the **zero-additional-cost** Kernel Inspired Router with Normalization ($\mathrm{KERN}$), an FFN-style router function, as an alternative to $\mathrm{Softmax}$. We demonstrate that this router generalizes both $\mathrm{Sigmoid}$- and $\mathrm{Softmax}$-based routers. **Based on empirical observations and established practices in FFN implementation, we recommend the use of $\mathrm{ReLU}$ activation and $\ell_2$-normalization in $\mathrm{KERN}$ router function.** Comprehensive experiments in MoE and LLM validate the effectiveness of the proposed FFN-style router function $\mathrm{KERN}$.

Continual learning (CL) enables models to acquire new tasks sequentially while retaining previously learned knowledge. However, most theoretical analyses focus on simplified, converged models or restrictive data distributions and therefore fail to capture how forgetting and generalization evolve during training in more general settings. Current theory faces two fundamental challenges: (i) analyses confined to the converged regime cannot characterize intermediate training dynamics; and (ii) establishing forgetting bounds requires two-sided bounds on the population risk for each task. To address these challenges, we analyze the training-time dynamics of forgetting and generalization in standard CL within the Neural Tangent Kernel (NTK) regime, showing that decreasing the loss’s Lipschitz constant and minimizing the cross-task kernel jointly reduce forgetting and improve generalization. Specifically, we (i) characterize intermediate training stages via kernel gradient flow and (ii) employ Rademacher complexity to derive both upper and lower bounds on population risk. Building on these insights, we propose \emph{OGD+}, which projects the current task’s gradient onto the orthogonal complement of the subspace spanned by gradients of the most recent task evaluated on all prior samples. We further introduce \emph{Orthogonal Penalized Gradient Descent} (OPGD), which augments OGD+ with gradient-norm penalization to jointly reduce forgetting and enhance generalization. Experiments on multiple benchmarks corroborate our theoretical predictions and demonstrate the effectiveness of OPGD, providing a principled pathway from theory to algorithm design in CL.

Hybrid Mamba-Transformer models have emerged as promising alternatives to pure Transformers, offering efficiency and competitive performance. However, they struggle to generalize beyond their training context windows, collapsing on long-context tasks. We provide the first systematic analysis of this failure, showing that it arises from uncontrolled state growth and uneven receptive field contributions across the hybrid architecture. Guided by this understanding, we introduce Universal Position Interpolation (UPI), a closed-form, training-free scaling method that unifies Mamba's cumulative decay with Transformer rotary frequency scaling. UPI selectively stabilizes unstable Mamba dynamics while rescaling Transformer encodings, controlling state growth and enabling reliable long-context generalization, with only a few auxiliary forward passes. Evaluation shows that UPI extends multiple state-of-the-art hybrid and pure Mamba models from 4K to up to 64K tokens on PG-19 perplexity, LongBench and RULER benchmarks, without sacrificing short-context accuracy. These findings establish the first principled bridge between Transformers and state-space models and open a new direction for training-free context extension methods for emerging hybrid models.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly advanced video understanding tasks, yet challenges remain in efficiently compressing visual tokens while preserving spatiotemporal interactions. Existing methods, such as LLaVA family, utilize simplistic pooling or interpolation techniques that overlook the intricate dynamics of visual tokens. To bridge this gap, we propose ST-GridPool, a novel training-free visual token enhancement method designed specifically for Video LLMs. Our approach integrates Pyramid Temporal Gridding (PTG), which captures multi-grained spatiotemporal interactions through hierarchical temporal gridding, and Norm-based Spatial Pooling (NSP), which preserves high-information visual regions by leveraging the correlation between token norms and semantic richness. Extensive experiments on various benchmarks demonstrate that ST-GridPool consistently enhances performance of Video LLMs without requiring costly retraining. Our method offers an efficient and plug-and-play solution for improving visual token representations. Our code is available in [https://anonymous.4open.science/r/ST-GridPool-85BE](https://anonymous.4open.science/r/ST-GridPool-85BE).

The past year has witnessed the significant advancement of video-based large language models. However, the challenge of developing a unified model for both short and long video understanding remains unresolved. Most existing video LLMs cannot handle hour-long videos, while methods custom for long videos tend to be ineffective for shorter videos and images. In this paper, we identify the key issue as the redundant content in videos. To address this, we propose a novel pooling strategy that simultaneously achieves token compression and instruction-aware visual feature aggregation. Our model is termed Prompt-guided Pooling LLaVA, or PPLLaVA for short. Specifically, PPLLaVA consists of three core components: the CLIP-based visual-prompt alignment that extracts visual information relevant to the user's instructions, the prompt-guided pooling that compresses the visual sequence to arbitrary scales using convolution-style pooling, and the clip context extension designed for lengthy prompt common in visual dialogue. Extensive experiments have validated the performance of our model. With superior throughput, PPLLaVA achieves better results on image benchmarks as a video LLM, while achieving state-of-the-art performance across various video benchmarks, excelling in tasks ranging from caption generation to multiple-choice questions, and handling video lengths from seconds to hours.

What happens when multiple compression methods are combined—does the order in which they are applied matter? Joint model compression has emerged as a powerful strategy to achieve higher efficiency by combining multiple methods such as pruning and quantization. A central but underexplored factor in joint model compression is the compression order, or the sequence of different methods within the compression pipeline. Most prior studies have either sidestepped the issue by assuming orthogonality between techniques, while a few have examined them only in highly constrained cases. Consequently, the broader role of compression order in shaping model performance remains poorly understood. In this paper, we address the overlooked problem of compression order and provide both theoretical and empirical analysis. We formulate the problem of optimizing the compression order and introduce the Progressive Intensity Hypothesis, which states that weaker perturbations should precede stronger ones. We provide theoretical guarantees showing that the relative benefit of one order increases with the underlying performance gap. Extensive experiments on both language and vision models validate the hypothesis, and further show its generality to broader setups such as multi-stage compression and mixed-precision quantization.

The "Reason-Then-Respond" paradigm, enhanced by Reinforcement Learning, has shown great promise in advancing Multimodal Large Language Models. However, its application to the video domain has led to specialized models that excel at either question answering (QA) or captioning tasks, but struggle to master both. Naively combining reward signals from these tasks results in mutual performance degradation, which we attribute to a conflict between their opposing task natures. To address this challenge, we propose a novel training framework built upon two intermediate proxy tasks: DarkEventInfer, which presents videos with masked event segments, requiring models to infer the obscured content based on contextual video cues; and MixVidQA, which presents interleaved video sequences composed of two distinct clips, challenging models to isolate and reason about one while disregarding the other. These proxy tasks compel the model to simultaneously develop both holistic, divergent understanding and precise, convergent reasoning capabilities. Embodying this framework, we present VidBridge-R1, the first versatile video reasoning model that effectively bridges the paradigm conflict. Extensive experiments show that VidBridge-R1 achieves significant performance gains on both QA and captioning within one model, demonstrating the efficacy of our approach in fostering more generalizable and powerful video understanding models. All code, models, and data will be made publicly available.

Accurate and efficient discrete video tokenization is essential for long video sequences processing. Yet, the inherent complexity and variable information density of videos present a significant bottleneck for current tokenizers, which rigidly compress all content at a fixed rate, leading to redundancy or information loss. Drawing inspiration from Shannon's information theory, this paper introduces \alg, a principled framework for adaptive video tokenization. We rigorously prove that existing data-agnostic training methods are suboptimal in representation length, and present a novel evidence lower bound (ELBO)-based algorithm that approaches theoretical optimality. Leveraging this framework, we develop a transformer-based adaptive compressor that enables adaptive tokenization. Empirical results demonstrate state-of-the-art compression performance, saving $20\%$ tokens without influence on performance, and achieving $2.3\times$ compression rates while still outperforming prior heuristic adaptive approaches. By allocating tokens according to informational richness, \alg enables a more compressed yet accurate tokenization for video representation, offering valuable insights for future research.

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called *subliminal learning*. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation—where the student only sees sampled tokens—raises a deeper question: *when and how does subliminal learning actually occur?* We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of *divergence tokens*—rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like paraphrasing prompts, are usually sufficient to suppress it.

Despite advancements in Multimodal Large Language Models (MLLMs), their proficiency in fine-grained video motion understanding remains critically limited. They often lack inter-frame differencing and tend to average or ignore subtle visual cues. Furthermore, while visual prompting has shown potential in static images, its application to videos' temporal complexities, particularly for fine-grained motion understanding, remains largely unexplored. We investigate whether inherent capability can be unlocked to boost MLLMs' motion perception and enable distinct visual signatures tailored to decouple object and camera motion cues. In this study, we introduce $\mathtt{MotionSight}$, a novel zero-shot method pioneering object-centric visual spotlight and motion blur as visual prompts to effectively improve fine-grained motion understanding without training. To convert this into valuable data assets, we curated $\mathtt{MotionVid-QA}$, the first large-scale dataset for fine-grained video motion understanding, with hierarchical annotations including SFT and preference data, $\Theta{(40K)}$ video clips and $\Theta{(87K)}$ QAs. Experiments show $\mathtt{MotionSight}$ achieves state-of-the-art open-source performance and competitiveness with commercial models. Using $\mathtt{MotionVid-QA}$, we fine-tuned $\mathtt{MotionChat}$ on Qwen2.5VL-7B, which attains 48.3\% overall accuracy on FAVOR-Bench that is comparable to Qwen2.5VL-72B's 48.1\%. In summary, we present a novel zero-shot method and a large-scale, high-quality dataset specifically for fine-grained motion understanding. All the code and annotations will be publicly available.