EgoMemReason is a new benchmark showing that even the best multimodal models achieve only 39.6% accuracy on reasoning tasks that require integrating sparse evidence across days in egocentric video.
hub Canonical reference
InternVideo2.5: Empowering Video MLLMs with Long and Rich Context Modeling
Canonical reference. 71% of citing Pith papers cite this work as background.
40
Pith papers citing it
Background
71% of classified citations
abstract
This paper aims to improve the performance of video multimodal large language models (MLLM) via long and rich context (LRC) modeling. As a result, we develop a new version of InternVideo2.5 with a focus on enhancing the original MLLMs' ability to perceive fine-grained details and capture long-form temporal structure in videos. Specifically, our approach incorporates dense vision task annotations into MLLMs using direct preference optimization and develops compact spatiotemporal representations through adaptive hierarchical token compression. Experimental results demonstrate this unique design of LRC greatly improves the results of video MLLM in mainstream video understanding benchmarks (short & long), enabling the MLLM to memorize significantly longer video inputs (at least 6x longer than the original), and master specialized vision capabilities like object tracking and segmentation. Our work highlights the importance of multimodal context richness (length and fineness) in empowering MLLM's innate abilites (focus and memory), providing new insights for future research on video MLLM. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2.5
hub tools
citation-role summary
background 5
method 1
other 1
citation-polarity summary
representative citing papers
LongEgoRefer is a new benchmark of 1,498 referring expressions in 45-minute average egocentric videos that exposes the failure of existing Video REC models on sparse long-form spatio-temporal grounding.
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.
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.
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.
SAVEMem improves streaming video understanding scores by adding semantic awareness to memory compression and query-adaptive retrieval without any model training.
A new benchmark converts video clips into shared grounded event records and tests models across physics, semantic, and control prompts under original, shuffled, ablated, and masked conditions, finding selective robustness and weak spatial performance.
InstAP introduces instance-aware pre-training with a new dual-granularity dataset InstVL that improves both fine-grained instance retrieval and global video understanding over standard VLP baselines.
Text-only contrastive fine-tuning of an MLLM with hard negatives produces embeddings that handle temporal, negation, and multimodal nuances in video retrieval and achieves SOTA performance.
Introduces TennisTV benchmark for evaluating 17 MLLMs on tennis video understanding from stroke-level to rally-level tasks with automated pipelines and human verification.
MotionAtlas supplies a 2,073-question benchmark, a self-bootstrap pipeline yielding 159k captions, and fine-tuned Video-MLLMs that deliver 5.2-point gains over Qwen3-VL-4B on motion tasks.
LVDR integrates keypoint-guided MCTS into a latent diffusion reasoning model to deliver competitive skill assessment accuracy alongside explicit visual reasoning trajectories on four sports and surgical datasets.
CoCoSI is a training-free multi-agent system for collaborative cognitive map construction that improves spatial understanding in arbitrary pretrained MLLMs.
LyraV uses FDTC and SToP for per-frame incremental decoding to reach 98.29% video synchrony at 3.89 FPS while preserving general understanding.
ViCuR introduces recoverable visual cues as teacher privilege in multimodal on-policy distillation, yielding +1.19 to +1.24 average gains over answer-based baselines across seven benchmarks with Qwen3-VL students.
GOPAgen proposes integrating video codec GOPs with a motion agent, GOP tree reasoning, structural memory, and motion vector database to improve efficiency and motion detail in agentic long-video VQA, reporting gains on MotionBench and EgoSchema.
IPIBench evaluates MLLMs on interactive proactive intelligence in streaming videos, identifies unstable triggering and poor coordination, and proposes the training-free IPI-Agent framework to improve performance across settings.
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.
LiteFrame is an efficient vision encoder backbone trained with Compressed Token Distillation and Language Model Adaptation to scale frame count in Video LLMs while cutting latency and raising accuracy.
Video-Zero is an annotation-free Questioner-Solver co-evolution framework that centers self-evolution on temporally localized evidence to improve video VLMs.
Video-LLMs fail physical reasoning due to semantic prior dominance rather than perception deficits; a new programmatic adversarial curriculum and visual-anchored reasoning chain enable substantial gains via standard LoRA fine-tuning.
A unified synthetic data generation pipeline produces unlimited annotated multimodal video data across multiple tasks, enabling models trained mostly on synthetic data to generalize effectively to real-world video understanding benchmarks.
DeSAP uses decoupled cross-modal similarity plus visual saliency to prune visual tokens in LVLMs, retaining 11.1% tokens for 10x FLOPs reduction and 98.1% performance on LLaVA-1.5-7B.
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.
citing papers explorer
-
EgoMemReason: A Memory-Driven Reasoning Benchmark for Long-Horizon Egocentric Video Understanding
EgoMemReason is a new benchmark showing that even the best multimodal models achieve only 39.6% accuracy on reasoning tasks that require integrating sparse evidence across days in egocentric video.
-
LongEgoRefer: A Benchmark for Long-Form Egocentric Video Referring Expression Comprehension
LongEgoRefer is a new benchmark of 1,498 referring expressions in 45-minute average egocentric videos that exposes the failure of existing Video REC models on sparse long-form spatio-temporal grounding.
-
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.
-
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.
-
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.
-
Semantic-Aware Adaptive Visual Memory for Streaming Video Understanding
SAVEMem improves streaming video understanding scores by adding semantic awareness to memory compression and query-adaptive retrieval without any model training.
-
Grounding Video Reasoning in Physical Signals
A new benchmark converts video clips into shared grounded event records and tests models across physics, semantic, and control prompts under original, shuffled, ablated, and masked conditions, finding selective robustness and weak spatial performance.
-
InstAP: Instance-Aware Vision-Language Pre-Train for Spatial-Temporal Understanding
InstAP introduces instance-aware pre-training with a new dual-granularity dataset InstVL that improves both fine-grained instance retrieval and global video understanding over standard VLP baselines.
-
Adapting MLLMs for Nuanced Video Retrieval
Text-only contrastive fine-tuning of an MLLM with hard negatives produces embeddings that handle temporal, negation, and multimodal nuances in video retrieval and achieves SOTA performance.
-
TennisTV: Do Multimodal Large Language Models Understand Tennis Rallies?
Introduces TennisTV benchmark for evaluating 17 MLLMs on tennis video understanding from stroke-level to rally-level tasks with automated pipelines and human verification.
-
MotionAtlas: Detailed Region Captioning for Motion-Centric Videos
MotionAtlas supplies a 2,073-question benchmark, a self-bootstrap pipeline yielding 159k captions, and fine-tuned Video-MLLMs that deliver 5.2-point gains over Qwen3-VL-4B on motion tasks.
-
Latent Visual Diffusion Reasoning with Monte Carlo Tree Search
LVDR integrates keypoint-guided MCTS into a latent diffusion reasoning model to deliver competitive skill assessment accuracy alongside explicit visual reasoning trajectories on four sports and surgical datasets.
-
CoCoSI: Collaborative Cognitive Map Construction for Spatial Intelligence
CoCoSI is a training-free multi-agent system for collaborative cognitive map construction that improves spatial understanding in arbitrary pretrained MLLMs.
-
Don't Pause: Streaming Video-Language Synchrony for Online Video Understanding
LyraV uses FDTC and SToP for per-frame incremental decoding to reach 98.29% video synchrony at 3.89 FPS while preserving general understanding.
-
ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation
ViCuR introduces recoverable visual cues as teacher privilege in multimodal on-policy distillation, yielding +1.19 to +1.24 average gains over answer-based baselines across seven benchmarks with Qwen3-VL students.
-
GOPAgen: Motion-Aware and Efficient Agentic Long-Video Understanding with Structural Memory and Hierarchical Reasoning
GOPAgen proposes integrating video codec GOPs with a motion agent, GOP tree reasoning, structural memory, and motion vector database to improve efficiency and motion detail in agentic long-video VQA, reporting gains on MotionBench and EgoSchema.
-
IPIBench: Evaluating Interactive Proactive Intelligence of MLLMs under Continuous Streams
IPIBench evaluates MLLMs on interactive proactive intelligence in streaming videos, identifies unstable triggering and poor coordination, and proposes the training-free IPI-Agent framework to improve performance across settings.
-
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.
-
LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs
LiteFrame is an efficient vision encoder backbone trained with Compressed Token Distillation and Language Model Adaptation to scale frame count in Video LLMs while cutting latency and raising accuracy.
-
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.
-
From Priors to Perception: Grounding Video-LLMs in Physical Reality
Video-LLMs fail physical reasoning due to semantic prior dominance rather than perception deficits; a new programmatic adversarial curriculum and visual-anchored reasoning chain enable substantial gains via standard LoRA fine-tuning.
-
All in One: A Unified Synthetic Data Pipeline for Multimodal Video Understanding
A unified synthetic data generation pipeline produces unlimited annotated multimodal video data across multiple tasks, enabling models trained mostly on synthetic data to generalize effectively to real-world video understanding benchmarks.
-
Decoupled Similarity for Task-Aware Token Pruning in Large Vision-Language Models
DeSAP uses decoupled cross-modal similarity plus visual saliency to prune visual tokens in LVLMs, retaining 11.1% tokens for 10x FLOPs reduction and 98.1% performance on LLaVA-1.5-7B.
-
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.
-
Streaming Video Instruction Tuning
Streamo is a streaming video LLM trained end-to-end on the new Streamo-Instruct-465K dataset that unifies multiple real-time video tasks with claimed strong temporal reasoning and generalization.
-
Perceive, Verify and Understand Long Video: Multi-Granular Perception and Active Verification via Interactive Agents
CogniGPT uses an interactive loop between a Multi-Granular Perception Agent and an Active Verification Agent to identify reliable clues in long videos with high accuracy and low frame usage.
-
SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics
SmolVLA is a small efficient VLA model that achieves performance comparable to 10x larger models while training on one GPU and deploying on consumer hardware via community data and chunked asynchronous action prediction.
-
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.
-
Bridging Video Understanding and Generation in a Unified Framework
Vega unifies video understanding and generation via shared vocabulary and hybrid autoregressive-diffusion architecture, reporting strong results on VBench and VideoMME.
-
LiveStarPro: Proactive Streaming Video Understanding with Hierarchical Memory for Long-Horizon Streams
LiveStarPro uses SVeD for response timing via perplexity, SCAM for incremental alignment, and TSHM for event-chain memory to achieve 28.9% better semantic correctness and 1.58x speedup on long video streams.
-
MultiToP: Learning to Patch Visual Tokens to Mitigate Hallucinations in Video Large Multimodal Models
MultiToP mitigates hallucinations in video multimodal models by training a Visual Token Patcher with information-guided rank calibration to selectively replace unreliable tokens, yielding 50.60% F1 gain on Vript-HAL and 18.58% accuracy gain on ActivityNet-QA.
-
Masked Diffusion Vision-Language Models for Temporal Action Localization
Adapts MDVLMs to TAL via planned training objective and step-level IoU reward, reporting gains over autoregressive baselines on ActivityNet and THUMOS datasets.
-
VidPrism: Heterogeneous Mixture of Experts for Image-to-Video Transfer
VidPrism introduces a heterogeneous temporal MoE with content-aware multi-rate sampling and bidirectional fusion for image-to-video transfer, claiming SOTA results on video benchmarks.
-
MuKV: Multi-Grained KV Cache Compression for Long Streaming Video Question-Answering
MuKV adds multi-grained KV cache compression at patch-frame-segment levels plus semi-hierarchical retrieval to raise accuracy and cut memory in long video question-answering.
-
VISD: Enhancing Video Reasoning via Structured Self-Distillation
VISD proposes structured self-distillation with a multi-dimensional judge model and direction-magnitude decoupling to improve token-level credit assignment and convergence speed in VideoLLM reasoning training.
-
High-Speed Vision Improves Zero-Shot Semantic Understanding of Human Actions
Higher temporal resolution in video significantly improves zero-shot semantic understanding of high-speed human actions like kendo.
-
How Should Video LLMs Output Time? An Analysis of Efficient Temporal Grounding Paradigms
A controlled study on compact video LLMs finds that continuous temporal decoding delivers the strongest accuracy-efficiency trade-off for video temporal grounding across three benchmarks.
-
OneThinker: All-in-one Reasoning Model for Image and Video
OneThinker unifies image and video reasoning in one model across 10 tasks via a 600k corpus, CoT-annotated SFT, and EMA-GRPO reinforcement learning, reporting strong results on 31 benchmarks plus some cross-task transfer.
-
VideoChat-R1: Enhancing Spatio-Temporal Perception via Reinforcement Fine-Tuning
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.
-
InternVideo3: Agentify Foundation Models with Multimodal Contextual Reasoning
InternVideo3 introduces Multimodal Contextual Reasoning and M^2LA attention to enable closed-loop evidence accumulation in long-video understanding and agentic tool use, reporting strong benchmark results.