ReToken: Improving Long-Context VLMs with Visual Retrieval Token

Name: ReToken: Improving Long-Context VLMs with Visual Retrieval Token
Uploaded: 2026-10-08T20:41:24.646706+00:00
Duration: 547 s
Description: RETOKEN introduces a single learnable embedding trained as an explicit retrieval target to select relevant visual tokens from a pre-filled cache. This method uses a two-pass strategy: first, retrieving query-relevant frames using the learned token, and second, answering only with those retrieved tokens. It significantly boosts performance on long video understanding and image QA benchmarks.

RETOKEN introduces a single learnable embedding trained as an explicit retrieval target to select relevant visual tokens from a pre-filled cache. This method uses a two-pass strategy: first, retrieving query-relevant frames using the learned token, and second, answering only with those retrieved tokens. It significantly boosts performance on long video understanding and image QA benchmarks.

Open the full episode