ResearchNovember 13, 2026
-42
🧊

V-JEPA-LM Crushes Open Video LLMs

A study examines whether refusal behavior can be uniformly removed across large language models such as Qwen, Gemma, Olmo, and Mistral using direction-based ablation.

#video models#V-JEPA#multimodal#benchmarks
V-JEPA-LM schlägt offene Video-LLMs
Share Article
🔥 What happened Researchers at Hanyang University dropped V-JEPA-LM, a video-language model that reads dynamics directly from video instead of relying on static, text-aligned features. Their 4B model scores 72.1 across eleven video benchmarks, beating the best open 4B (61.1) and even the best open 8B (64.8). 💡 Why it matters Standard VLMs answer speed questions at chance level, while a linear probe on frozen V-JEPA features hits 87%. Counterfactual supervision adds 4.5 points on average and a massive 38 points on speed questions—using just 4.53M samples and 32 tokens per frame, versus multi-stage pipelines with up to 256 tokens. ⚡ Our take This is a wake-up call for anyone scaling video LLMs with more text and tokens. The encoder decides what's readable, not the data volume. Stick with SigLIP-style encoders and you'll keep missing the dynamics.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.