REVIEW 16 cited by
InternVideo2: Scaling Foundation Models for Multimodal Video Understanding
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/.
Forward citations
Cited by 16 Pith papers
-
MentalThink: Shaping Thoughts in Mental SVG World
MLLMs that generate and render SVG sketches as multi-turn intermediate reasoning steps reach 55.1% on VSIBench and 76.0% on MindCube, far above the Qwen2.5-VL-7B backbone.
-
AdsQA: Towards Advertisement Video Understanding
AdsQA adds an ad-video question-answering benchmark and ReAd-R, a GRPO-trained model that beats 7B baselines but not larger closed models.
-
HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly
A dual-branch video classifier uses depth and spatiotemporal features plus rank-weighted losses to categorize human-centric AI forgeries into spatial, appearance, and motion anomaly types on a new auto-labeled benchmark.
-
LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs
A training-free token compression method using semantic connected components in space and time keeps video understanding accuracy high even when retaining only 5-10% of visual tokens.
-
DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification
DejaVid replaces mean-pooling over clips with a learned, time-weighted dynamic-time-warping distance to class centroids, adding 0.5-0.7% top-1 accuracy on three video benchmarks.
-
VideoMolmo: Spatio-Temporal Grounding Meets Pointing
A video language model that conditions each frame on earlier frames via a temporal attention module, predicts text-requested object points, and uses SAM2-based bidirectional mask fusion to outperform prior models on v...
-
AV-Reasoner: Improving and Benchmarking Clue-Grounded Audio-Visual Counting for MLLMs
A clue-grounded audio-visual counting benchmark over 497 long videos and an RL-trained counting model, whose headline result is undermined by training on the DVD-Counting evaluation benchmark.
-
HuMoCon: Concept Discovery for Human Motion Understanding
A framework that combines explicit video-motion feature alignment with velocity-aware masked autoencoding to improve LLM-based human motion and video question answering.
-
RTime-QA: A Benchmark for Atomic Temporal Event Understanding in Large Multi-modal Models
RTime-QA is a video-question benchmark where models choose between temporally opposite descriptions of the same event, and current AI models score far below humans.
-
Temporal Object Captioning for Street Scene Videos from LiDAR Tracks
A rule-based pipeline converts LiDAR tracks into template captions of traffic dynamics, and training SwinBERT on them lowers the Visual Bias Measure across three datasets.
-
Whom to Respond To? A Transformer-Based Model for Multi-Party Social Robot Interaction
A multi-task transformer with two KL-divergence losses improves a social robot's when-and-whom-to-respond accuracy on a new multi-party HRI dataset.
-
How Far Can Off-the-Shelf Multimodal Large Language Models Go in Online Episodic Memory Question Answering?
A training-free pipeline that summarizes egocentric video clips into a few kilobytes of text per minute and answers multiple-choice episodic memory questions with an LLM reasoner reaches 56.0% accuracy on QAEgo4D-Clos...
-
EVA02-AT: Egocentric Video-Language Understanding with Spatial-Temporal Rotary Positional Embeddings and Symmetric Optimization
EVA02-AT combines full-dimension spatial and temporal rotary position embeddings with a symmetric multi-similarity loss to improve egocentric video-text retrieval.
-
An Empirical study on LLM-based Log Retrieval for Software Engineering Metadata Management
A natural-language log retrieval pipeline using LLM-generated video and signal descriptions shows that prompt and model choice matter, but its proposed reliability metrics are not validated against any ground truth.
-
HCQA-1.5 @ Ego4D EgoSchema Challenge 2025
An ensemble of LLMs with confidence filtering and low-confidence re-reasoning reaches 77% accuracy on the EgoSchema benchmark, up from 75% for the prior HCQA system.
-
Reinforcement Learning: From Algorithms To Foundation Models
A dissertation uniting the author's published results: non-exploitable Nash-DQN policies and the FightLadder benchmark for games, plus diffusion/consistency-model world models for RL — a compilation rather than new results.
Discussion (0). Sign in to comment.