Skip to content

Publications / PRPO

PRPO: Collaborative Online Policy Learning in Personalized RLHF

A. Vladimirov, E. Gurov, V. Smirnov, V. Solodkin, A. Beznosikov

Preprint

OpenReview ↗Figure

Expected validation win rate on heterogeneous TLDR with GRPO: the two PR (PRPO) curves stay above isolated and FedAvg training across hyperparameter assignments. — Personalized RLHF on heterogeneous TLDR (GRPO): PRPO (PR) sustains a higher expected validation win rate than isolated and FedAvg training as more clients collaborate. Open full-resolution figure (PDF) ↗

A detailed write-up of this work is coming soon.

← Back to home