V-JEPA-LM: Counterfactual Supervision Teaches LLMs to Understand Video

V-JEPA-LM (1B-8B) couples a frozen V-JEPA 2.1 video encoder with an LLM and trains with counterfactual supervision on temporally reversed videos, lifting 4B-level accuracy from 61.1 to 72.1.

2026-11-16 ~ 2026-11-17 · 2 related posts