EarlyTom: Early Token Compression Completes Fast Video Understanding

Researchers introduced EarlyTom, a training-free token compression framework for video LLMs.

Unlike post-encoder pruning, it compresses visual tokens inside the vision encoder with decoupled spatial token selection.

On LLaVA-OneVision-7B, it cuts time-to-first-token by up to 2.65x and FLOPs by up to 61%.

Accuracy stays close to the full-token baseline, making video understanding models faster and cheaper to deploy.