VideoLVR: Latent Visual Reasoning over Grounded Video Evidence
VideoLVR lets models revisit relevant video details during reasoning without substantially expanding their context.
The method builds latent visual evidence from local regions at relevant moments, interleaving those visual states with text reasoning. A training objective teaches the model to reconstruct visual tokens using a vocabulary derived from the input video. The authors also extend reinforcement learning to optimize visual and text token choices together using final-answer rewards. They report improvements across multiple video benchmarks, especially tasks requiring fine-grained visual evidence; the abstract provides no numerical results.
ArXiv · AI/CL/LG's note
The method builds latent visual evidence from local regions at relevant moments, interleaving those visual states with text reasoning. A training objective teaches the model to reconstruct visual tokens using a vocabulary derived from the input video. The authors also extend reinforcement learning to optimize visual and text token choices together using final-answer rewards. They report improvements across multiple video benchmarks, especially tasks requiring fine-grained visual evidence; the abstract provides no numerical results.
ArXiv · AI/CL/LG's note
score 5