Video-Language Models

Video Large Language Models (Video-LLMs) achieve strong performance in video understanding, but their excessive visual tokens bring substantial computational overhead. Existing training-free compression methods improve inference efficiency by reducing visual tokens, yet they often rely on local adjacent-frame similarity for temporal redundancy estimation or allocate token budgets mainly according to segment length. Such designs are sensitive to frame-level noise and fail to capture the non-uniform information distribution of real-world videos. To address these challenges, we propose InfoMerge, a training-free visual token compression method that improves token utilization through robust redundancy estimation and content-aware budget allocation. Specifically, we propose the Temporal Fingerprint Difference: a segment-level second-order temporal redundancy estimation strategy, which models the temporal similarity structure of tokens at the same spatial positions within each segment. We further introduce Content-Aware Budget Allocation (CABA), which dynamically allocates segment-level token budgets based on segment uniqueness and spectral-entropy-based representational richness. By reducing repeated preservation of redundant static regions and allocating more tokens to informative segments, InfoMerge makes better use of the limited token budget while maintaining strong performance. Extensive experiments show that InfoMerge achieves strong efficiency--accuracy trade-offs across multiple benchmarks and backbones, with more pronounced advantages under aggressive compression. On LLaVA-OneVision-7B, InfoMerge retains 98.8% of the original average performance while reducing 85% of visual tokens and achieving a 4.24-fold speedup in the prefill stage.

As multimodal large language models (MLLMs) have advanced to process video inputs, concerns have emerged about their potential for malicious misuse. Prior jailbreak studies have shown that safety alignment in MLLMs can be bypassed through visual inputs, yet it remains unclear which properties of video inputs induce this vulnerability. To address this gap, we introduce Multi-Clip Video (MCV) SafetyBench, a dataset of 2,920 videos designed to evaluate how the diversity of video inputs affects the vulnerability of MLLMs. Each video consists of multiple short clips depicting diverse contexts related to a harmful query. Experiments on eight representative video MLLMs show that attack success consistently increases with the number of clips. Our results further indicate that the video modality is (1) more vulnerable than the image modality, (2) more vulnerable to dynamic videos than to static videos, and (3) more vulnerable when videos contain more diverse contexts. Building on these findings, we propose a defense strategy that leverages the relative robustness of the image modality.

Video understanding is shifting from the offline paradigm -- taking a fully recorded video as input and producing a single answer after it ends -- toward real-time interaction, in which the model perceives new frames while still replying, revises its answer as new evidence appears, and remains silent when there is nothing to say. We present MOSS-Video-Preview to validate this paradigm. Our central claim is that perception must not be blocked by generation; its natural realization is a two-channel architecture. We argue that a cross-attention backbone is better suited to real-time vision-language fusion than the prevailing decoder-only design: visual features enter through a side channel rather than joining the autoregressive sequence, so perception and generation run on separate, non-blocking pathways -- reducing the frequency of visual processing and exposing a clean channel-wise interface for independent compression. We complement this with a data synthesis pipeline that converts dense captions into real-time understanding QA whose answers are revised to match what the model has perceived so far, and we specialize an offline model on these data to elicit real-time behavior. Our model trails the strong Qwen2.5-VL-7B baseline overall -- a gap we attribute primarily to data and scale rather than the architecture -- yet attains competitive offline video and multimodal understanding, remains robust on the spatial and fine-grained temporal reasoning central to real-time use, and acquires behaviors that offline models lack: continuous perception, answer revision, and timely silence. On a single H200 with 256 frames per video, it achieves about a 5x speedup in time to first token and 2.7x higher decoding throughput, with negligible degradation in offline ability. Our study of paradigm, architecture, and data outlines a viable path toward real-time video understanding.

This study introduces an intriguing phenomenon in Video LLMs: rather than merely translating frames into textual embeddings, Video LLMs establish a continuous manifold, token interface, allowing visual tokens to operate as standalone entities within the architecture. Exploiting this discovery, we propose V-LynX, a scalable framework that integrates novel modalities into Video LLMs by repurposing the internalized interface. Departing from conventional paradigms that necessitate heavy modality-specific encoders or paired supervision, V-LynX employs a lightweight auxiliary pathway in parallel with the frozen vision encoder. Our method integrates new sensory inputs with intrinsic video priors by aligning both attention responses and statistical distributions using unpaired unimodal data sets. This ensures manifold compatibility while preserving the integrity of the Video LLMs. Extensive benchmarks demonstrate that V-LynX achieves SOTA and efficiency across audio-visual QA, 3D reasoning, high-frame-rate, and multi-view video understanding. The code is available at https://github.com/park-jungin/lynx.

Accurately predicting future events is fundamental to content understanding and decision-making across various domains. While prior research has primarily focused on text or short-video scenarios, long-video event prediction, characterized by vast multimodal context and more complex narratives, remains underexplored. Meanwhile, although recent Long-Video Language Models (LVLMs), built on Large Language Models (LLMs) and Vision-Language Models (VLMs), have shown promise in long-video question answering and summarization, they struggle to generalize to event prediction, as they can neither precisely extract event-related details nor perform fine-grained analysis of event development. To address this gap, we propose VISTA, a multi-level event semantics mining framework for long-video event prediction. Initially, VISTA applies a character-centric visual prompt to precisely extract event-related visual details, enhancing detail-level semantics; subsequently, it employs a knowledge-enhanced iterative retrieval strategy, guiding the LLM to progressively construct logically coherent event chains, thereby improving event-level narratives; ultimately, VISTA adopts a human-like propose-then-retrieve strategy to generate diverse future-oriented proposals and integrate multi-level clues, producing robust and accurate predictions. Extensive experiments on real-world datasets validate the effectiveness of VISTA for long-video event prediction.

In the era of Large Video-Language Models (LVLMs), the computational necessity of sparse frame sampling creates a fundamental ``temporal gap'', rendering models blind to critical causal transitions. Existing solutions relying on generative hallucination (e.g., latent diffusion) or autoregressive extrapolation often fail to maintain semantic consistency over long horizons, suffering from object vanishing and energetic instability. We propose a paradigm shift from probabilistic generation to variational mechanics with the \textbf{Semantic Least Action Principle (SLAP)}. Drawing a rigorous isomorphism between classical mechanics and semantic dynamics, we model the latent video trajectory as a path on a Riemannian manifold governed by a Semantic Lagrangian. By formulating the interpolation task as a Boundary Value Problem (BVP) solved via the discrete Euler-Lagrange equations, SLAP naturally enforces object persistence without pixel-level rendering. Extensive experiments show the effectiveness of our proposed SLAP.

Temporal action localization (TAL) requires recognizing the target event and localizing its start and end times precisely in untrimmed videos. Recent vision-language formulations improve semantic reasoning and support language-conditioned outputs, but their autoregressive decoders still generate tokens from left to right, preventing later semantic evidence from revising earlier timestamp predictions. We adapt masked diffusion vision-language models (MDVLMs) to TAL so that semantic tokens and boundary tokens remain editable throughout iterative denoising with bidirectional attention, allowing temporal boundaries and semantic content to be refined jointly. Direct adaptation, however, creates two TAL-specific mismatches: standard masked diffusion training corrupts all positions uniformly at random, but the time tokens are more reliable when enough semantic context is available; and token-level cross-entropy does not reflect temporal IoU. To address these mismatches, we introduce a Planned Training Objective that uses boundary-aware masking and step-weighted reconstruction to rehearse the late recovery of time tokens, together with a Step-Level IoU Reward that provides overlap-aware supervision during denoising. A standard sequence-level cross-entropy term provides the base reconstruction signal. Experiments on ActivityNet-RTL, ActivityNet-1.3, and THUMOS-14 show that MDVLM-TAL improves both temporal reasoning and boundary localization over autoregressive vision-language baselines, with especially strong gains under stricter temporal IoU criteria.

Current vision-language models (VLMs) typically stitch together separate image encoders and language decoders via multi-stage alignment, a modular framework that inevitably fragments pixel-level signals across frames and scatters early pixel-word interactions. In parallel, native VLMs, despite impressive performance on single images, remain largely unexplored in multi-image, video understanding, and spatial intelligence. Hence, we introduce NEO-ov, a native foundation model that learns cross-frame and pixel-word correspondence end-to-end, without any external encoders, auxiliary adapters, or post-hoc fusion. By eliminating module boundaries entirely, NEO-ov enables fine-grained and unified spatiotemporal modeling to emerge natively inside the model. Notably, NEO-ov largely narrows the gap to modular counterparts while excelling at fine-grained visual perception, validating that native "one-vision" architectures are not only feasible but competitive at scale. Beyond empirical performance, we unveil systematic architectural analyses and detailed training recipes to facilitate subsequent native multimodal modeling. Our code and models are publicly available at: https://github.com/EvolvingLMMs-Lab/NEO.

Driven by the wave of large language models, Video-Language Models (VLMs) have become a significant yet challenging technology to bridge the gap between videos and texts. Although previous VLM works have made significant progress, almost all of them implicitly assume that all the texts are predefined by the specific template. In real-world applications, such a strict assumption is impossible to satisfy since 1) predefining all the texts is extremely time-consuming and labor-intensive. 2) these predefined text inputs are too restrictive and user-unfriendly, limiting their applications. It is observed that given a video input, texts with similar semantics but different templates lead to various performances. To this end, in this paper, we propose a novel plug-and-play framework for various VLM-based methods to fully bridge videos and texts. Specifically, we first generate positive and negative texts from the original ones to target specific text components. Then, we propose an attribute-based text reasoning strategy to mine fine-grained textual semantics of generated texts. Finally, we utilize videos as guidance to conduct cross-modal bridging by designing a self-weighted loss. Extensive experiments show that the proposed method can serve as the plug-and-play module to effectively improve the performance of state-of-the-art VLMs.

Video-Language Models (VLMs) have demonstrated impressive multi-modal reasoning capabilities across diverse computer vision applications. However, these VLMs are task-specific and assume that both video and language inputs are complete. However, real-world VLM applications might face challenges due to deactivated sensors (e.g., cameras are unavailable due to data privacy), yielding modality-incomplete data and leading to inconsistency between training and testing data. While straightforward incomplete input can boast training generalization-ability and lead to training failure, its potential risks to VLMs regarding safety and trustworthiness have been largely neglected. To this end, we make the first attempt to propose a unified incomplete video-language model to process the incomplete multi-modal inputs. Extensive experimental results show that our method can serve as a plug-and-play module for previous works to improve their performance in various multi-modal tasks.

A key capability for video understanding is reliably linking subjects to events across time, yet whether Video Large Language Models (VideoLLMs) actually achieve this remains unclear. In this work, we introduce DistractionBench to evaluate whether VideoLLMs can robustly link subjects and events in the presence of unrelated video segments. Through controlled interventions, such as inserting short advertisement clips into longer videos, we show that VideoLLMs frequently hallucinate interactions between entities from different segments, incorrectly attributing actions from injected advertisements to subjects in the main video. We characterize this systematic hallucination as bag-of-events (BoE) behavior, where models process videos as collections of events rather than temporally structured sequences. Evaluating 11 popular VideoLLMs, we find that all models exhibit substantial BoE behavior. Our findings suggest that VideoLLMs lack reliable mechanisms for temporal grounding and motivate the development of models with more robust subject-event association.

Existing video captioning methods struggle to balance visual fidelity and redundancy: holistic captions are compact but lose fine-grained evidence, whereas segment-wise captions improve coverage but introduce heavy redundancy. We propose CodecCap, a codec-inspired framework for high-fidelity dense video captioning. Analogous to video codecs, CodecCap represents videos using keyframe and residual captions. Keyframe captions exhaustively encode stable visual context, while residual captions capture temporally only localized actions, motions and changes. This effectively preserves fine-grained visual evidence while reducing redundant descriptions. To quantify the fidelity of captions, we introduce VidCapQA, a caption-then-QA benchmark with 1,000 questions across 14 capability dimensions. Results on VidCapQA show that captions directly generated by strong VLMs still miss many visual details, highlighting caption representation as a critical bottleneck. Experiments show that CodecCap significantly surpasses direct captioning with the same underlying VLMs, suggesting keyframe-residual captioning a way for high-fidelity video-language supervision. We further use CodecCap to construct CodecVDC-100K, a large-scale dense captioning dataset with anchor, residual, scene-level, and video-level supervision.

Many video reasoning tasks require tracking motion, temporal order, and evolving visual states across frames. Existing methods built on large vision-language models (LVLMs) often address this challenge by externalizing reasoning through textual chain-of-thought (CoT), keyframe selection, repeated frame reinsertion, or external tool use. While effective, such pipelines increase inference-time latency and engineering complexity, and they force temporal-visual evidence to be serialized into text or repeatedly re-encoded from frames. Inspired by the intuition that visual reasoning can occur implicitly before verbalization, we propose STORMS (Spatial-Temporal reasOning via inteRnalized Modeling), a two-stage framework that teaches LVLMs to reason through bounded continuous latent trajectories instead of explicit textual CoT. In Stage I, STORMS aligns latent tokens with thought-video representations derived from generated videos, grounding the latent states in dynamic visual evidence. In Stage II, the model is further trained with answer-only supervision, encouraging the reasoning process to be internalized without step-by-step annotations. Generated thought videos are used only during training; at inference, STORMS performs a bounded latent rollout without regenerating videos, reinserting frames, or invoking external visual tools. Experiments on VideoMME, MVBench, TempCompass, and MMVU show that STORMS improves video reasoning accuracy while substantially reducing inference overhead compared with tool or video-generation-based reasoning pipelines.

Video Large Language Models (Video-LLMs) have made rapid progress on temporal video understanding, yet many fail at a basic perceptual primitive: signed image-plane motion direction. On simple videos of a single object moving left, right, up, or down, most Video-LLMs perform near chance, with above-chance cases largely attributable to prediction biases rather than genuine direction understanding. We call this failure directional motion blindness. We localize the failure by tracing motion direction information through the Video-LLM pipeline. Motion direction remains linearly accessible from the vision encoder, projector, and LLM hidden states, but the readout fails to bind this signal to the correct verbal answer option, revealing a direction binding gap. Although synthetic motion direction instruction tuning reduces this gap on the source domain, motion direction concept vector analysis shows that visual complexity weakens the signal magnitude and limits out-of-domain generalization. We introduce MoDirect, a dataset family for motion direction instruction tuning and evaluation, and DeltaDirect, a diagnosis-driven, projector-level objective that predicts normalized 2-D motion vectors from adjacent-frame feature deltas. On MoDirect-SynBench, instruction tuning with DeltaDirect improves motion direction accuracy from 25.9% to 85.9%. On MODIRECT-REALBENCH, DeltaDirect improves realworld motion direction accuracy by 21.4 points over the vanilla baseline without real-world tuning data, while preserving standard video-understanding performance. Our project page is available at https://jong980812.github.io/which-way-did-it-move/

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://github.com/bingjunluo/ST-GridPool.

Video large language models (Video LLMs) can achieve strong video-QA accuracy without reliably tracking spatiotemporal dynamics. A model may answer a motion question from static cues, for example, and give the same prediction even after the underlying motion is reversed. Correctness-based reinforcement learning does not directly address this problem because it rewards the final answer without requiring the policy to respond to task-relevant video dynamics. We propose Counterfactual Relational Policy Optimization (CRPO), which explicitly trains Video LLMs to respond to controlled changes in the visual input. For each training example, CRPO constructs a counterfactual video, such as a horizontally flipped or temporally reversed version of the original, and jointly optimizes rollouts from both videos under a shared policy. Factual supervision anchors what the model should answer, while counterfactual supervision constrains when that answer should change: predictions should change when an intervention alters task-relevant dynamics and remain stable when the queried property is preserved. This coupling provides a direct behavioral learning signal without requiring ground-truth labels for transformed videos or annotated spatiotemporal reasoning traces, while discouraging indiscriminate answer changes. To evaluate this property, we introduce DyBench, a paired counterfactual benchmark with 3{,}014 videos and a strict pair-accuracy metric. Across paired spatiotemporal evaluations and standard video benchmarks, CRPO improves sensitivity to motion and temporal changes while improving performance on general video understanding. The gains also extend to segment reordering, a transformation never used during training, suggesting that CRPO learns sensitivity to video dynamics beyond the training interventions. The project website can be found at https://ddz16.github.io/crpo.github.io/ .

Current Video-LLM approaches for Video Temporal Grounding (VTG) typically rely on direct timestamp generation from an unstructured visual-token stream, often leading to brittle numerics and inconsistent boundaries. To address this, we propose Foresee-to-Ground (F2G), a framework that reformulates VTG as a verifiable Identify-then-Measure problem. F2G integrates Predictive Temporal Perception with Evidence-Driven Reasoning: it learns boundary-sensitive temporal representations to build a video-wide evidence pool of candidate event segments, and exposes these segments to the LLM as citable evidence units that bind boundary prediction to explicit event hypotheses. By decoupling event identification from precise boundary measurement, F2G stabilizes grounding and makes predictions verifiable. Extensive experiments demonstrate that F2G consistently improves grounding accuracy across diverse benchmarks, transfers robustly across different Video-LLM backbones, and preserves general video understanding capabilities.

Existing video-language models can generate factual descriptions of road events but lack control over how these events are expressed: their tone, urgency, or style. This limits deployment in communication-critical settings where the effectiveness of a message depends on both content and presentation, not just factual accuracy. To mitigate this, we introduce a comprehensive dataset-model-evaluation suite for tone-controllable road video captioning. Our human-validated data generation pipeline expands road-video corpora with diverse tonal annotations and multi-tone captions, yielding the RoadTones-51K dataset. We propose RoadTones-VL-CoT, a controllable video-to-text model that also generates tone-conditioned Chain-of-Thought intermediate drafts for interpretability. We also introduce RoadTones-Eval, a new evaluation suite that jointly measures factual consistency and tone adherence. In addition, we conducted a user study whose results validate caption quality, tone control, and factual consistency. Together, these contributions lay the foundation for context-sensitive tone-controllable video captioning.

Recent advances in Video Large Language Models (Video-LLMs) have greatly expanded multimodal reasoning capabilities. However, the massive number of visual tokens extracted from long video sequences incurs prohibitive computational costs, limiting their deployment in real-world scenarios. Existing training-free token compression methods select tokens based on attention magnitude as a proxy for semantic importance, but often overlook positional bias and rely only on short-term temporal locality, leading to redundant spatio-temporal coverage and inefficient token usage. We present DynaTok, a training-free, temporally adaptive and bias-aware token compression framework that allocates token budgets across both temporal and spatial dimensions. Through a lightweight exponential moving average (EMA) memory, the Temporal Budget Allocation (TBA) module dynamically assigns fewer tokens to redundant frames and more to novel frames, capturing long-term temporal variation. The Spatial Budget Allocation (SBA) module complements this by selecting spatially diverse and semantically important features using activation-based attention maps, while leveraging a spatial memory to reduce redundancy from previously selected regions and mitigate positional bias. DynaTok integrates seamlessly with existing Video-LLMs such as LLaVA-OneVision and LLaVA-Video without retraining, and effectively preserves semantic coverage under aggressive compression. Experiments on four representative VideoQA benchmarks-MVBench, LongVideoBench, MLVU, and VideoMME-show that DynaTok retains over 95% of baseline accuracy even with a 90% token reduction, surpassing recent training-free approaches. These results demonstrate that DynaTok provides a principled foundation for efficient and robust video reasoning, paving the way toward real-time streaming video understanding with future Video-LLMs.

Children acquire language grounding with remarkable robustness from limited visuo-linguistic input in ways that surpass today's best large multimodal models. Recent research suggests current vision-language models (VLMs) trained on curated web data fail to generalize to the sparse, weakly-aligned egocentric streams produced by wearable devices, embodied agents, and infant head-cams -- and no fixed evaluation pipeline exists for measuring progress on this regime. We train VLMs on datasets with varying degrees of semantic alignment between visual and linguistic inputs, including naturalistic infant and adult egocentric videos, and evaluate them with a comprehensive suite spanning multimodal language grounding and unimodal vision and language tasks. At the core of this suite is Machine-DevBench, a corpus-grounded benchmark of lexical and grammatical competence, automatically generated from the model's training vocabulary across logarithmic frequency bins to eliminate the train/eval mismatch and low statistical power of prior developmental benchmarks. Our results show that current VLM paradigms hinge on the tight semantic alignment of curated data and fail to exploit the weakly-aligned signal that dominates naturalistic egocentric input -- the very regime in which humans thrive. To motivate progress, we introduce the EgoBabyVLM Challenge to drive the development of models capable of grounded language learning from the kind of naturalistic data that human infants experience.

The fundamental challenge in scaling Video Large Language Models (Video LLMs) to long-form video lies in managing the explosion of visual-token context length. Existing strategies predominantly focus on "post-hoc" token reduction -- reducing visual tokens after feature extraction to alleviate the LLM's computational overhead. While these methods effectively reduce the number of visual tokens, we observe that the primary latency bottleneck then shifts from the LLM to the expensive per-frame processing of the vision encoder. To address this, we introduce LiteFrame, a strong, yet highly efficient video encoder backbone for Video LLMs. To train LiteFrame, we propose Compressed Token Distillation (CTD), a novel training framework that teaches a compact student vision encoder to directly predict information-dense, spatio-temporally compressed representations produced by a large teacher vision model, effectively bypassing redundant computation. When coupled with further Language Model Adaptation (LMA), this approach results in a new latency-accuracy Pareto frontier -- compared with InternVL3-8B, LiteFrame provides a 35% reduction in end-to-end latency while processing 8$\times$ more frames and improves average video understanding accuracy across multiple benchmarks. Our results demonstrate a new potential path to unlocking longer-form video understanding under fixed compute budgets.

Despite rapid progress in video-capable MLLMs, we find that their apparent audio understanding in videos is often vision-driven: models rely on visual cues to infer or hallucinate acoustic information, rather than verifying the audio stream. This issue appears across both state-of-the-art open-source omni models and leading closed-source models from providers such as Google and OpenAI. We characterize this failure mode as an audio-visual Clever Hans effect, in which models appear (falsely) audio-grounded, but actually exploit visual-acoustic correlations without verifying whether the audio and visual streams are truly aligned. To systematically study this behavior, we introduce Thud, an intervention-driven probing framework based on three counterfactual audio edits: Shift, which tests temporal synchronization; Mute, which tests sound existence; and Swap, which tests audio-visual consistency. Beyond diagnosis, we further study a two-stage alignment recipe: intervention-derived preference pairs teach audio verification, while event-level general video preferences regularize the model against over-specialization. Our best 10K-sample recipe improves average performance across the three intervention dimensions by 28 percentage points, while slightly improving performance on general video and audio-visual QA benchmarks.

Multimodal video summarization requires visual features that align semantically with language generation. Traditional approaches rely on CNN features trained for object classification, which represent visual concepts as discrete categories not aligned with natural language. We propose ClipSum, a framework that leverages frozen CLIP vision-language features with explicit temporal modeling and dimension-adaptive fusion for instructional video summarization. CLIP's contrastive pre-training on 400M image-text pairs yields visual features semantically aligned with the linguistic concepts that text decoders generate, bridging the vision-language gap at the representation level. On YouCook2, ClipSum achieves 33.0% ROUGE-1 versus 30.5% for ResNet-152 with 4x lower dimensionality (512 vs. 2048), demonstrating that semantic alignment matters more than feature capacity. Frozen CLIP (33.0%) surpasses fine-tuned CLIP (32.3%), showing that preserving pre-trained alignment is more valuable than task-specific adaptation. https://github.com/aqeeelmirza/clipsum

As Video Large Language Models (Video-LLMs) scale to longer and more complex videos, their inference cost grows rapidly due to the large volume of visual tokens accumulated across frames. Training-free token compression has emerged as a practical solution to this bottleneck. However, existing temporal compression methods rely primarily on cross-frame token similarity or segmentation heuristics, overlooking each token's semantic role within its frame and failing to adapt compression strength to the compressibility of each frame pair. In this work, we propose OTT-Vid, a transport-derived allocation framework for temporal token compression. Our approach consists of two stages: spatial pruning identifies representative content within each frame, and optimal transport (OT) is then solved between neighboring frames to estimate temporal compressibility. We formulate this OT with non-uniform token mass, which protects semantically important tokens from aggressive compression, and a locality-aware cost that captures both feature and spatial disparities. The resulting transport plan jointly balances token importance and matching cost, while its total cost defines the transport difficulty of each frame pair, which we use to allocate compression budgets dynamically. Experiments on six benchmarks spanning video question answering and temporal grounding show that OTT-Vid preserves 95.8% of VQA and 73.9% of VTG performance while retaining only 10% of tokens, consistently outperforming existing state-of-the-art training-free compression methods.

Streaming video understanding demands more than watching longer videos: assistants must decide when to speak in real time, balancing responsiveness against verbosity. Yet most video-language models (VideoLLMs) are trained for offline inference, and existing streaming benchmarks externalize this timing decision to the evaluator. We address this gap with RealStreamEval, a frame-level multi-turn evaluation protocol that exposes models to sequential observations and penalizes unnecessary responses. Under this protocol, we observed that strong offline VideoLLMs retain useful visual understanding but lack an interaction policy for deciding when to respond. Motivated by this observation, we propose EvoStreaming, a self-evolved streaming adaptation framework in which the base model itself acts as data generator, relevance annotator, and roll-out policy to synthesize streaming trajectories without external supervision. With only $1{,}000$ self-generated samples ($139\times$ less than the leading streaming instruction-tuning approach) and no architectural changes, EvoStreaming consistently improves the overall RealStreamEval score by up to $10.8$ points across five open VideoLLM backbones (Qwen2/2.5/3-VL, InternVL-3.5, MiniCPM-V4.5) while largely preserving offline video performance. These results suggest that data-efficient interaction tuning is a practical path for adapting existing VideoLLMs to streaming assistants.

Video large language models (Video-LLMs) have made strong progress in general video understanding, but their ability to maintain temporal object consistency remains underexplored. Existing benchmarks often emphasize event recognition, action understanding, or coarse temporal reasoning, while rarely testing whether models can preserve the identity, state, and continuity of the same object across occlusion, disappearance, reappearance, state transitions, and cross-object interactions. We introduce TOC-Bench, a diagnostic benchmark for evaluating temporal object consistency in Video-LLMs. TOC-Bench is object-track grounded: each queried subject is linked to a per-frame trajectory and a structured temporal event timeline. To ensure that questions require temporally ordered visual evidence rather than language priors, single-frame shortcuts, or unordered frame cues, we design a three-layer temporal-necessity filtering protocol, which removes 60.7% of candidate QA pairs and retains 17,900 temporally dependent items across 10 diagnostic dimensions. From this pool, we construct a human-verified benchmark with 2,323 high-quality QA pairs over 1,951 videos. Experiments on representative Video-LLMs show that temporal object consistency remains a major unsolved challenge, with notable weaknesses in event counting, event ordering, identity-sensitive reasoning, and hallucination-aware verification, even when models perform well on general video understanding benchmarks. These results suggest that object-centric temporal coherence is a key bottleneck for current Video-LLMs, and that TOC-Bench provides a focused platform for diagnosing and improving object-aware temporal reasoning. The resource is available at https://github.com/cjzcjz666/toc_bench.git.

The Arrow-of-Time (AoT) task, determining whether a video plays forward or backward by recognizing temporal irreversibility, is one humans solve with near-perfect accuracy, yet frontier Video Large Language Models (Video-LLMs) perform only modestly above chance. This gap raises a key question: do visual backbones fail to encode temporal information, or does information bottleneck lie elsewhere in the Video-LLM architecture? We address this question by isolating the vision encoder from the Video-LLM and tracing temporal information across the encoder, projector, and LLM. We find that video-centric encoders with explicit temporal modeling encode strong temporal signals, whereas frame-centric encoders do not. However, when video-centric representations are passed through a standard Video-LLM architecture, performance often collapses, revealing a bottleneck of temporal information flow. We identify projector design as a key factor: Q-Former disrupts temporal information, while a time-preserved MLP projection substantially improves the LLM's access to such information. Our layer-wise analysis further shows temporal representation dynamics across encoder layers. Guided by these findings, we build a Video-LLM with temporal-aware video-centric encoder, time-preserved projector, and AoT supervision, surpassing human performance on AoT$_{PPB}$ with 98.1% accuracy, and improving broader temporal reasoning tasks by up to 6.0 points on VITATECS-Direction and 1.3 points on TVBench. Our results show that temporal reasoning in Video-LLMs requires both effective temporal encoding and reliable transfer of this information to the LLM.

Video-language models (VLMs) face rapid inference costs as visual token counts scale with video length. For example, 32 frames at $448{\times}448$ resolution already yield >8,000 visual tokens in Qwen3-VL, making LLM prefill the dominant throughput bottleneck. Existing methods often rely on global similarity or attention-guided compression, incurring offsets to their gains. We propose \textbf{Temporal Token Fusion (TTF)}, a training-free, plug-and-play pre-LLM token compression framework that exploits structured temporal redundancy in video. TTF automatically selects an anchor frame, then for each subsequent frame, performs a local window similarity search (e.g.,$3\times 3$), fusing tokens that exceed a threshold. The compressed sequence maintains positional consistency across both prefill and decoding through coordinate realignment, enabling seamless integration with existing VLM pipelines. On Qwen3-VL-8B with threshold t=0.70, TTF removes about 67% of visual tokens while retaining 99.5% of the baseline accuracy and introducing only ${\approx}0.16$,GFLOPs of matching overhead. Overall, TTF offers a practical, efficient solution for video understanding. The code is available at https://github.com/Cominder/ttf

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 rollout level advantages computed from 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 curriculum scheduling and EMA based teacher stabilization to support robust optimization over long video sequences. Experiments on diverse benchmarks show that VISD consistently outperforms strong baselines, improving answer accuracy and spatio temporal grounding quality. Notably, VISD reaches these gains with nearly 2x faster convergence in optimization steps, highlighting the effectiveness of structured self supervision in improving both performance and sample efficiency for VideoLLMs.

Text-based role-playing models can imitate character styles, but often fail to capture scene atmosphere and evolving tension, which are crucial for immersive applications such as VR games and interactive narratives. We study video-grounded role-playing dialogue and introduce EBM-RL (Eye--Brain--Mouth Reinforcement Learning), a decoupled GRPO-based framework that separates observation (<perception>), reasoning (<think>), and utterance generation (<answer>). This design mimics the human See-Think-Speak process, enabling the model to ground dialogue in visual perception before reasoning and response generation. To optimize this See-Think-Speak process, EBM-RL integrates complementary rewards for scene--text alignment, perceptual--cognitive utility, answer faithfulness, and format consistency. Extensive experiments show that EBM-RL substantially outperforms text-only role-playing baselines and larger-scale vision-language models on our immersive role-playing benchmark, improving both visual-atmosphere consistency and character authenticity. Moreover, EBM-RL demonstrates strong zero-shot transfer to out-of-domain VideoQA benchmarks without additional fine-tuning. We also release an open-source dataset for video-grounded role-playing dialogue.