EgoSound is a new benchmark with 7315 QA pairs across seven tasks to evaluate egocentric sound understanding in multimodal large language models.
hub Baseline reference
MLVU: Benchmarking Multi-task Long Video Understanding
Baseline reference. 62% of citing Pith papers use this work as a benchmark or comparison.
62
Pith papers citing it
4
external citations · Pith
Baseline
62% of classified citations
abstract
The evaluation of Long Video Understanding (LVU) performance poses an important but challenging research problem. Despite previous efforts, the existing video understanding benchmarks are severely constrained by several issues, especially the insufficient lengths of videos, a lack of diversity in video types and evaluation tasks, and the inappropriateness for evaluating LVU performances. To address the above problems, we propose a new benchmark called MLVU (Multi-task Long Video Understanding Benchmark) for the comprehensive and in-depth evaluation of LVU. MLVU presents the following critical values: \textit{1)} The substantial and flexible extension of video lengths, which enables the benchmark to evaluate LVU performance across a wide range of durations. \textit{2)} The inclusion of various video genres, e.g., movies, surveillance footage, egocentric videos, cartoons, game videos, etc., which reflects the models' LVU performances in different scenarios. \textit{3)} The development of diversified evaluation tasks, which enables a comprehensive examination of MLLMs' key abilities in long-video understanding. The empirical study with 23 latest MLLMs reveals significant room for improvement in today's technique, as all existing methods struggle with most of the evaluation tasks and exhibit severe performance degradation when handling longer videos. Additionally, it suggests that factors such as context length, image-understanding ability, and the choice of LLM backbone can play critical roles in future advancements. We anticipate that MLVU will advance the research of long video understanding by providing a comprehensive and in-depth analysis of MLLMs.
hub tools
citation-role summary
dataset 11
background 5
citation-polarity summary
representative citing papers
Future-L1 interleaves latent visual spans with text in MLLM decoding, trained on a custom Future-L1-50K dataset via LA-DAPO RL, and reports SOTA gains on FutureBench (61.0 to 85.4) and TwiFF-Bench (2.44 to 3.04).
VideoKR supplies 315K knowledge-intensive video reasoning examples and a dedicated benchmark, with experiments indicating post-training gains on reasoning tasks that require both video content and external knowledge.
VSTAT benchmark shows state-of-the-art MLLMs perform far below humans and only modestly above answer-prior baselines on visual state tracking, failing at visual perception despite correct textual reasoning.
SVI-Bench provides 35K hours of sports video with 9 tasks across four cognitive levels, revealing models drop from ~74% on action QA to 5% on agentic evidence integration.
R3-Streaming uses cascaded control with age-aware memory forgetting and TB-GRPO reinforcement learning to reach SOTA scores of 57.92 on OVO-Bench and 76.36 on StreamingBench with 95-96% fewer visual tokens.
MemLens benchmark shows long-context LVLMs lose accuracy with length while memory agents lose visual fidelity, with multi-session reasoning below 30% for most systems and neither approach solving the task alone.
CoRDS selects a compact KV-cache subset via joint-space coreset coverage and log-det diversity to outperform token-wise heuristics on long-video VLM benchmarks.
SYNCR benchmark shows leading MLLMs reach only 52.5% average accuracy on cross-video reasoning tasks against an 89.5% human baseline, with major weaknesses in physical and spatial reasoning.
StoryTR is a new benchmark and agentic data pipeline that adds explicit Theory of Mind reasoning chains to train smaller video retrieval models, yielding a 15% relative IoU gain over larger baselines on narrative content.
Mosaic uses cross-modal clusters as the unit for KVCache organization in VLMs to achieve up to 1.38x speedup in streaming long-video understanding.
AdaSpark delivers up to 57% FLOP reduction in Video-LLMs for long videos through adaptive cube- and token-level sparsity without apparent loss in performance on hour-scale benchmarks.
SVAgent improves long video question answering by constructing storylines via multi-agent collaboration and aligning cross-modal predictions for more robust, human-like reasoning.
SIV-Bench is a new video benchmark with 2,792 clips and 5,455 QA pairs that evaluates MLLMs on social scene understanding, state reasoning, and dynamics prediction using social relation theory.
UnifiedReward is the first unified reward model that jointly assesses multimodal understanding and generation to provide better preference signals for aligning vision models via DPO.
WorldSense provides the first benchmark requiring synergistic audio-video-text understanding on 1,662 real-world videos and 3,172 QA pairs, where the best current multimodal LLM reaches only 65.1% accuracy.
Video-MMMU benchmark shows large multimodal models exhibit steep performance drops on higher cognitive tasks when learning from professional videos and lag significantly behind humans in knowledge acquisition.
Audit of four VideoQA benchmarks reveals text-only shortcuts in VLMs; new diagnostics Blind Gap, Visual Gain, and Shortcut Score quantify and filter visual dependence.
HPP decouples perception from reasoning in long-video VLMs by having an LLM run iterative programmatic probes on hierarchically segmented video, reporting gains on LongVideoBench, EgoSchema, VideoMME, and MLVU.
Fits a model where logit-accuracy scales linearly in log frame budget B with distance-dependent exponent α(D) that decays log-linearly with temporal distance D, based on 155k binary predictions across ten models.
Presents Streaming-Train-248K dataset, Streaming Harness system, and Streaming-Eval benchmark to enable VLMs for proactive, memory-equipped streaming video understanding.
MemoryCard organizes long videos into self-contained topic-aware Memory Cards that improve long-video QA accuracy by up to 21.8% relative under fixed visual-token budgets.
RLVR exhibits correct-set turnover where solved problems regress during training, and a periodic review mechanism exploiting a repair-window principle improves retention and performance over baselines.
MOSS-Video-Preview introduces a cross-attention architecture and synthesized real-time QA data to enable continuous perception, answer revision, and faster inference in video-language models compared to decoder-only designs.
citing papers explorer
-
EgoSound: Benchmarking Sound Understanding in Egocentric Videos
EgoSound is a new benchmark with 7315 QA pairs across seven tasks to evaluate egocentric sound understanding in multimodal large language models.
-
Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction
Future-L1 interleaves latent visual spans with text in MLLM decoding, trained on a custom Future-L1-50K dataset via LA-DAPO RL, and reports SOTA gains on FutureBench (61.0 to 85.4) and TwiFF-Bench (2.44 to 3.04).
-
VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding
VideoKR supplies 315K knowledge-intensive video reasoning examples and a dedicated benchmark, with experiments indicating post-training gains on reasoning tasks that require both video content and external knowledge.
-
Benchmarking Visual State Tracking in Multimodal Video Understanding
VSTAT benchmark shows state-of-the-art MLLMs perform far below humans and only modestly above answer-prior baselines on visual state tracking, failing at visual perception despite correct textual reasoning.
-
SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence
SVI-Bench provides 35K hours of sports video with 9 tasks across four cognitive levels, revealing models drop from ~74% on action QA to 5% on agentic evidence integration.
-
An Efficient Streaming Video Understanding Framework with Agentic Control
R3-Streaming uses cascaded control with age-aware memory forgetting and TB-GRPO reinforcement learning to reach SOTA scores of 57.92 on OVO-Bench and 76.36 on StreamingBench with 95-96% fewer visual tokens.
-
MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models
MemLens benchmark shows long-context LVLMs lose accuracy with length while memory agents lose visual fidelity, with multi-session reasoning below 30% for most systems and neither approach solving the task alone.
-
CoRDS: Coreset-based Representative and Diverse Selection for Streaming Video Understanding
CoRDS selects a compact KV-cache subset via joint-space coreset coverage and log-det diversity to outperform token-wise heuristics on long-video VLM benchmarks.
-
SYNCR: A Cross-Video Reasoning Benchmark with Synthetic Grounding
SYNCR benchmark shows leading MLLMs reach only 52.5% average accuracy on cross-video reasoning tasks against an 89.5% human baseline, with major weaknesses in physical and spatial reasoning.
-
StoryTR: Narrative-Centric Video Temporal Retrieval with Theory of Mind Reasoning
StoryTR is a new benchmark and agentic data pipeline that adds explicit Theory of Mind reasoning chains to train smaller video retrieval models, yielding a 15% relative IoU gain over larger baselines on narrative content.
-
Mosaic: Cross-Modal Clustering for Efficient Video Understanding
Mosaic uses cross-modal clusters as the unit for KVCache organization in VLMs to achieve up to 1.38x speedup in streaming long-video understanding.
-
AdaSpark: Adaptive Sparsity for Efficient Long-Video Understanding
AdaSpark delivers up to 57% FLOP reduction in Video-LLMs for long videos through adaptive cube- and token-level sparsity without apparent loss in performance on hour-scale benchmarks.
-
SVAgent: Storyline-Guided Long Video Understanding via Cross-Modal Multi-Agent Collaboration
SVAgent improves long video question answering by constructing storylines via multi-agent collaboration and aligning cross-modal predictions for more robust, human-like reasoning.
-
SIV-Bench: A Video Benchmark for Social Interaction Understanding and Reasoning
SIV-Bench is a new video benchmark with 2,792 clips and 5,455 QA pairs that evaluates MLLMs on social scene understanding, state reasoning, and dynamics prediction using social relation theory.
-
Unified Reward Model for Multimodal Understanding and Generation
UnifiedReward is the first unified reward model that jointly assesses multimodal understanding and generation to provide better preference signals for aligning vision models via DPO.
-
WorldSense: Evaluating Real-world Omnimodal Understanding for Multimodal LLMs
WorldSense provides the first benchmark requiring synergistic audio-video-text understanding on 1,662 real-world videos and 3,172 QA pairs, where the best current multimodal LLM reaches only 65.1% accuracy.
-
Video-MMMU: Evaluating Knowledge Acquisition from Multi-Discipline Professional Videos
Video-MMMU benchmark shows large multimodal models exhibit steep performance drops on higher cognitive tasks when learning from professional videos and lag significantly behind humans in knowledge acquisition.
-
From Accuracy to Visual Dependence: Auditing and Filtering Modality Collapse in Traffic VideoQA
Audit of four VideoQA benchmarks reveals text-only shortcuts in VLMs; new diagnostics Blind Gap, Visual Gain, and Shortcut Score quantify and filter visual dependence.
-
HPP: Hierarchical Programmatic Probing for Long Video Understanding by Decoupling Perception and Reasoning
HPP decouples perception from reasoning in long-video VLMs by having an LLM run iterative programmatic probes on hierarchically segmented video, reporting gains on LongVideoBench, EgoSchema, VideoMME, and MLVU.
-
How Well Can Your Video Model Remember? Measuring Memory-Budget Trade-offs in Long Video Understanding
Fits a model where logit-accuracy scales linearly in log frame budget B with distance-dependent exponent α(D) that decays log-linearly with temporal distance D, based on 155k binary predictions across ten models.
-
Harnessing Streaming Video in the Wild
Presents Streaming-Train-248K dataset, Streaming Harness system, and Streaming-Eval benchmark to enable VLMs for proactive, memory-equipped streaming video understanding.
-
MemoryCard: Topic-Aware Multi-Modal Clue Compression for Long-Video Question Answering
MemoryCard organizes long videos into self-contained topic-aware Memory Cards that improve long-video QA accuracy by up to 21.8% relative under fixed visual-token budgets.
-
Learning to Solve, Forgetting to Retain: Correct-Set Turnover in RLVR
RLVR exhibits correct-set turnover where solved problems regress during training, and a periodic review mechanism exploiting a repair-window principle improves retention and performance over baselines.
-
MOSS-Video-Preview: Toward Real-Time Video Understanding via Cross-Attention
MOSS-Video-Preview introduces a cross-attention architecture and synthesized real-time QA data to enable continuous perception, answer revision, and faster inference in video-language models compared to decoder-only designs.
-
TeachObs: A Human-Validated Benchmark for Multimodal Teaching Observation and Model Evaluation
TeachObs provides human-validated segment- and lesson-level labels on 30 classroom videos and finds frontier VLMs inconsistent, frame-inflated, and over-generous versus expert raters.
-
DynFrame: Adaptive Reasoning-Driven Multimodal Framework with Dynamic Frame Augmentation for Complex Video Understanding
DynFrame introduces tokenized learnable span-density retrieval and Segment-Decoupled GRPO in video MLLMs, achieving competitive or SOTA results on six benchmarks with 4B and 8B models.
-
OProver: A Unified Framework for Agentic Formal Theorem Proving
OProver-32B achieves top Pass@32 scores on MiniF2F, ProverBench, and PutnamBench by combining continued pretraining with iterative agentic proving, retrieval, SFT on repairs, and RL on unresolved cases using a 6.86M-proof dataset.
-
Video-Zero: Self-Evolution Video Understanding
Video-Zero is an annotation-free Questioner-Solver co-evolution framework that centers self-evolution on temporally localized evidence to improve video VLMs.
-
Video Active Perception: Effective Inference-Time Long-Form Video Understanding with Vision-Language Models
VAP is a training-free active-perception method that improves zero-shot long-form video QA performance and frame efficiency up to 5.6x in VLMs by selecting keyframes that differ from priors generated by a text-conditioned video model.
-
One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding
XComp reaches extreme video compression (one token per selective frame) via learnable progressive token compression and question-conditioned frame selection, lifting LVBench accuracy from 42.9 percent to 46.2 percent after tuning on 2.5 percent of standard data.
-
ForestPrune: High-ratio Visual Token Compression for Video Multimodal Large Language Models via Spatial-Temporal Forest Modeling
ForestPrune prunes 90% of visual tokens in video MLLMs like LLaVA-OneVision while retaining 95.8% accuracy by modeling tokens as spatial-temporal forests and scoring importance via tree depth and node roles.
-
Towards Effective Long Video Understanding of Multimodal Large Language Models via One-shot Clip Retrieval
OneClip-RAG enables MLLMs to handle long videos via one-shot clip retrieval and unified chunking-retrieval, delivering performance gains like matching GPT-5 level on MLVU with high efficiency on standard GPUs.
-
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
InternVL3.5 advances open-source multimodal models with Cascade RL for +16% reasoning gains and ViR for 4x inference speedup, with the 241B model reaching SOTA among open-source MLLMs on multimodal, reasoning, and agentic tasks.
-
ReGATE: Learning Faster and Better with Fewer Tokens in MLLMs
ReGATE introduces a teacher-student adaptive token elision method that reduces training tokens to 38% while matching or exceeding baseline accuracy on multimodal benchmarks.
-
One Trajectory, One Token: Grounded Video Tokenization via Panoptic Sub-object Trajectory
TrajViT tokenizes videos via panoptic sub-object trajectories, achieving 10x token reduction and outperforming ViT3D by 6% on retrieval and 5.2% on VideoQA tasks with faster training and inference.
-
LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning
LLaDA-V is a diffusion-based multimodal large language model that reaches competitive or state-of-the-art results on visual instruction tasks while using a non-autoregressive architecture.
-
LiveVLM: Efficient Online Video Understanding via Streaming-Oriented KV Cache and Retrieval
LiveVLM introduces VSB and PaR to compress and retrieve KV cache in streaming video LLMs, enabling LLaVA-OneVision to reach SOTA accuracy among training-free query-agnostic and training-based online models.
-
InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models
InternVL3-78B sets a new open-source SOTA of 72.2 on MMMU via native joint multimodal pre-training, V2PE, MPO, and test-time scaling while remaining competitive with proprietary models.
-
SmolVLM: Redefining small and efficient multimodal models
SmolVLM-256M outperforms a 300-times larger model using under 1 GB GPU memory, while the 2.2B version matches state-of-the-art VLMs at half the memory cost.
-
MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models
MotionBench is a new benchmark showing poor fine-grained motion understanding in VLMs and proposes TE Fusion to improve performance with higher frame rates.
-
VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling
VideoChat-Flash applies hierarchical video token compression to achieve ~50x reduction in context length for long videos while maintaining near-original performance on long-context benchmarks.
-
LLaVA-Video: Video Instruction Tuning With Synthetic Data
LLaVA-Video-178K is a new synthetic video instruction dataset that, when combined with existing data to train LLaVA-Video, produces strong results on video understanding benchmarks.
-
TuringViT: Making SOTA Vision Transformers Accessible to All
TuringViT claims a new ViT design with linear attention and curated data that matches SOTA performance using 10% of typical pretraining data while supporting dynamic resolutions and improving VLM integration.
-
Towards Fast and Effective Long Video Understanding of Multimodal Large Language Models via Adaptive Quasi-Gaussian Sampling
AdaQ is a training-free adaptive quasi-Gaussian sampling method for keyframe selection that improves long-video understanding in MLLMs and can outperform GPT-4o with 64 frames.
-
Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models
GeoVR distills camera pose, depth, scale, and multi-scale 3D features from pre-trained models into MLLMs via video supervision to improve spatial reasoning.
-
Efficient Long-Context Modeling in Diffusion Language Models via Block Approximate Sparse Attention
BA-Att introduces pre-downsampled block selection with norm-sorting and diagonal covariance correction to approximate sparse attention, yielding up to 6.95x speedup at 50% sparsity across language, multimodal, and video models.
-
EgoCoT-Bench: Benchmarking Grounded and Verifiable Operation-Centric Chain of Thought Reasoning for MLLMs
EgoCoT-Bench provides 3,172 verifiable QA pairs across perception, anticipation, and reasoning tasks on egocentric videos, revealing that many MLLMs give answer-correct but evidence-inconsistent explanations.
-
What Limits Vision-and-Language Navigation ?
StereoNav reaches new benchmark highs on R2R-CE and RxR-CE and improves real-robot reliability by supplying persistent target-location priors and stereo-derived geometry that stay stable under lighting changes and blur.
-
TTF: Temporal Token Fusion for Efficient Video-Language Model
TTF fuses temporally redundant visual tokens via local similarity search in a plug-and-play way, cutting ~67% tokens on Qwen3-VL-8B while retaining 99.5% accuracy with minimal overhead.
-
Qwen2.5-VL Technical Report
Qwen2.5-VL reports a vision-language model family using native dynamic-resolution ViT and absolute time encoding that matches GPT-4o on document and diagram tasks while supporting hour-long videos with second-level localization.