MedHorizon benchmark reveals current multimodal LLMs achieve only 41.1% accuracy on long medical videos due to failures in sparse evidence retrieval and procedural reasoning.
hub Canonical reference
VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling
Canonical reference. 71% of citing Pith papers cite this work as background.
26
Pith papers citing it
Background
71% of classified citations
abstract
Long-context video modeling is critical for multimodal large language models (MLLMs), enabling them to process movies, online video streams, and so on. Despite its advances, handling long videos remains challenging due to the difficulty in efficiently understanding the extremely long video context. This paper aims to address this issue from aspects of model architecture, training data, training strategy and evaluation benchmark. First, we propose a novel Hierarchical video token Compression (HiCo) method, which leverages visual redundancy in long videos to compress long video context from Clip-level to Video-level, reducing the computation significantly while preserving essential details, achieving an extreme compression ratio of approximately 1/50 with almost no performance loss. Second, we introduce a multi-stage short-to-long learning scheme, a large-scale dataset of real-world long videos named LongVid, and a challenging ``Multi-Hop Needle-In-A-Video-Haystack'' benchmark. Finally, we build a powerful video MLLM named VideoChat-Flash, which shows a leading performance on both mainstream long and short video benchmarks at the 2B and 7B model scale. It first gets 99.1% accuracy over 10,000 frames in NIAH among open-source models.
hub tools
citation-role summary
background 5
baseline 1
method 1
citation-polarity summary
representative citing papers
Molmo2 delivers state-of-the-art open-weight video VLMs with new grounding datasets and training methods that outperform prior open models and match or exceed some proprietary ones on pointing and tracking tasks.
MLLMs drop from over 85% accuracy on action presence to under 50% on matched action-denial videos, exposing a causal verification gap that causal graph prompts partially close.
A proposer-solver agent pair achieves supervised-level video temporal grounding and fine-grained captioning from 2.5K unlabeled videos via self-reinforcing evolution.
StreamPro introduces a benchmark and training method using CB-Stream Loss and GRPO to enable proactive decision-making in streaming videos, achieving 41.5 on StreamPro-Bench compared to 10.4 previously.
TOC-Bench is a new diagnostic benchmark that reveals major weaknesses in temporal object consistency for Video-LLMs, including event counting, ordering, identity reasoning, and hallucination avoidance.
OmniVTG creates a new large-scale open-world VTG dataset using iterative concept-gap filling and timestamped captioning, paired with a three-stage self-correction CoT paradigm that yields SOTA zero-shot results on four existing benchmarks.
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.
LongVideo-R1 trains a reasoning agent on 33K trajectories to intelligently select informative video clips via iterative refinement and RL, achieving better accuracy-efficiency tradeoffs on long video QA benchmarks.
VideoThinker uses LLM-generated synthetic tool trajectories in caption space grounded to video frames to train agentic VideoLLMs that outperform baselines on long-video benchmarks.
VISE is the first benchmark for sycophancy in Video-LLMs, with two training-free mitigation strategies based on key-frame selection and internal representation steering.
AdaCodec introduces a predictive visual code that cuts visual token use in video MLLMs by sending full frames only on high predictive cost and otherwise encoding inter-frame changes as P-tokens, yielding better benchmark scores at lower budgets.
Introduces VIP identification task, releases Temporal-VIP dataset, and presents VIP-Net framework that achieves 67.3% accuracy on identifying important persons in videos while providing rationale similarity of 0.63.
MLLMs know event timing during prefill via sparse Temporal Grounding Heads but lose it in autoregressive decoding; restricting visual context to the high-attention interval at inference time improves VTG performance on three benchmarks.
VisMMoE exploits visual-expert affinity via token pruning to achieve up to 2.68x faster VL-MoE inference on memory-constrained hardware while keeping accuracy competitive.
Tuna-2 shows that direct pixel embeddings can replace vision encoders in unified multimodal models, achieving competitive generation and stronger understanding at scale.
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.
POINTS-Long is a dual-mode multimodal large language model that uses dynamic visual token scaling to retain 97.7-99.7% accuracy on long-form tasks with 1/40 to 1/10th the tokens and supports streaming via detachable KV-cache.
Tempo uses a 6B SVLM as a local temporal compressor with training-free adaptive token allocation to achieve SOTA long-video understanding at 0.5-16 tokens per frame, scoring 52.3 on 4101s LVBench under 8K budget.
G2F-RAG converts retrieved knowledge subgraphs into a single visual reasoning frame appended to videos, enabling training-free and interpretable improvements for LMM-based video reasoning on knowledge-intensive tasks.
Cambrian-S introduces VSI-SUPER benchmarks for long-horizon spatial recall and counting, shows data scaling yields 30% gains on existing tests, and demonstrates a self-supervised next-latent predictor using surprise outperforms baselines on the new spatial supersensing tasks.
StateKV is an inference-time technique that replaces quadratic self-attention prefill in video VLMs with a fixed-capacity importance-based recurrent state, keeping accuracy near full attention on long-video benchmarks without retraining.
PyraVid is a hierarchical multimodal memory system that structures long videos into pyramids to improve long-horizon reasoning and evidence aggregation.
Reinforcement fine-tuning with temporal rewards produces VideoChat-R1, a video MLLM showing large gains on spatio-temporal perception benchmarks such as +31.8 temporal grounding and +31.2 object tracking.
citing papers explorer
-
MedHorizon: Towards Long-context Medical Video Understanding in the Wild
MedHorizon benchmark reveals current multimodal LLMs achieve only 41.1% accuracy on long medical videos due to failures in sparse evidence retrieval and procedural reasoning.
-
Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding
Molmo2 delivers state-of-the-art open-weight video VLMs with new grounding datasets and training methods that outperform prior open models and match or exceed some proprietary ones on pointing and tracking tasks.
-
Learning to Deny: Action Denial in Multimodal Large Language Models
MLLMs drop from over 85% accuracy on action presence to under 50% on matched action-denial videos, exposing a causal verification gap that causal graph prompts partially close.
-
EvoGround: Self-Evolving Video Agents for Video Temporal Grounding
A proposer-solver agent pair achieves supervised-level video temporal grounding and fine-grained captioning from 2.5K unlabeled videos via self-reinforcing evolution.
-
StreamPro: From Reactive Perception to Proactive Decision-Making in Streaming Video
StreamPro introduces a benchmark and training method using CB-Stream Loss and GRPO to enable proactive decision-making in streaming videos, achieving 41.5 on StreamPro-Bench compared to 10.4 previously.
-
TOC-Bench: A Temporal Object Consistency Benchmark for Video Large Language Models
TOC-Bench is a new diagnostic benchmark that reveals major weaknesses in temporal object consistency for Video-LLMs, including event counting, ordering, identity reasoning, and hallucination avoidance.
-
OmniVTG: A Large-Scale Dataset and Training Paradigm for Open-World Video Temporal Grounding
OmniVTG creates a new large-scale open-world VTG dataset using iterative concept-gap filling and timestamped captioning, paired with a three-stage self-correction CoT paradigm that yields SOTA zero-shot results on four existing benchmarks.
-
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.
-
LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding
LongVideo-R1 trains a reasoning agent on 33K trajectories to intelligently select informative video clips via iterative refinement and RL, achieving better accuracy-efficiency tradeoffs on long video QA benchmarks.
-
VideoThinker: Building Agentic VideoLLMs with LLM-Guided Tool Reasoning
VideoThinker uses LLM-generated synthetic tool trajectories in caption space grounded to video frames to train agentic VideoLLMs that outperform baselines on long-video benchmarks.
-
AdaCodec: A Predictive Visual Code for Video MLLMs
AdaCodec introduces a predictive visual code that cuts visual token use in video MLLMs by sending full frames only on high predictive cost and otherwise encoding inter-frame changes as P-tokens, yielding better benchmark scores at lower budgets.
-
Mining Multi-Modality Spatio-Temporal Cues for Video Important Person Identification
Introduces VIP identification task, releases Temporal-VIP dataset, and presents VIP-Net framework that achieves 67.3% accuracy on identifying important persons in videos while providing rationale similarity of 0.63.
-
MLLMs Know When Before Speaking: Revealing and Recovering Temporal Grounding via Attention Cues
MLLMs know event timing during prefill via sparse Temporal Grounding Heads but lose it in autoregressive decoding; restricting visual context to the high-attention interval at inference time improves VTG performance on three benchmarks.
-
VisMMOE: Exploiting Visual-Expert Affinity for Efficient Visual-Language MoE Offloading
VisMMoE exploits visual-expert affinity via token pruning to achieve up to 2.68x faster VL-MoE inference on memory-constrained hardware while keeping accuracy competitive.
-
Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation
Tuna-2 shows that direct pixel embeddings can replace vision encoders in unified multimodal models, achieving competitive generation and stronger understanding at scale.
-
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.
-
POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs
POINTS-Long is a dual-mode multimodal large language model that uses dynamic visual token scaling to retain 97.7-99.7% accuracy on long-form tasks with 1/40 to 1/10th the tokens and supports streaming via detachable KV-cache.
-
Small Vision-Language Models are Smart Compressors for Long Video Understanding
Tempo uses a 6B SVLM as a local temporal compressor with training-free adaptive token allocation to achieve SOTA long-video understanding at 0.5-16 tokens per frame, scoring 52.3 on 4101s LVBench under 8K budget.
-
Graph-to-Frame RAG: Visual-Space Knowledge Fusion for Training-Free and Auditable Video Reasoning
G2F-RAG converts retrieved knowledge subgraphs into a single visual reasoning frame appended to videos, enabling training-free and interpretable improvements for LMM-based video reasoning on knowledge-intensive tasks.
-
Linear Scaling Video VLMs for Long Video Understanding
StateKV is an inference-time technique that replaces quadratic self-attention prefill in video VLMs with a fixed-capacity importance-based recurrent state, keeping accuracy near full attention on long-video benchmarks without retraining.
-
PyraVid: Hierarchical Multimodal Memory for Long-Horizon Video Reasoning
PyraVid is a hierarchical multimodal memory system that structures long videos into pyramids to improve long-horizon reasoning and evidence aggregation.
-
LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence
LLaVA-OV-2 uses codec-stream tokenization and a shared 3D RoPE to improve video, spatial, and tracking performance over Qwen3-VL-8B, while introducing the JumpScore benchmark for fine-grained motion localization.