Variance-aware RLHF adds a variance-weighted KL penalty to PPO and reduces reward variance and the risk of underperforming the reference policy in the paper's experiments.
The claude 3 model family: Opus, sonnet, haiku
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Towards Reliable, Uncertainty-Aware Alignment
Variance-aware RLHF adds a variance-weighted KL penalty to PPO and reduces reward variance and the risk of underperforming the reference policy in the paper's experiments.