Using predicted human gaze to shape rewards lets RLHF converge 1.3 to 2 times faster with similar final policy quality.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Enhancing RLHF with Human Gaze Modeling
Using predicted human gaze to shape rewards lets RLHF converge 1.3 to 2 times faster with similar final policy quality.