SFO: Piloting VLM Feedback for Offline RL

Jacob Beck · 2025

Read SFO: Piloting VLM Feedback for Offline RL by Jacob Beck (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-01062.txt.

Description

Jacob Beck feeds vision model grades on short clips into offline learning, in a paper from two thousand twenty-five. Web-scale vision models judge success well but cannot steer acts. Whole-run grades tangle the stitching that offline learning needs.

Short sub-runs fix the length trap, and clip grades judge growth without act labels. A plain filtered clone method beat fancier rivals in the tests. A backward filter drops clips that lead into failure, for calm learning. The set of tricks is named SFO with a clone form called SFBC. Readers get AI feedback that respects how offline learning breaks.

Catalog record

CREATORJacob Beck
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-01062

Browse the full library · About the Tower of Babel