Scalable Reinforcement Post-Training Beyond Static Human Prompts: Evolving Alignment via Asymmetric Self-Play

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Ye, Ziyu, Agarwal, Rishabh, Liu, Tianqi, Joshi, Rishabh, Velury, Sarmishta, Le, Quoc V., Tan, Qijun, Liu, Yuan
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866916682135502848
author Ye, Ziyu
Agarwal, Rishabh
Liu, Tianqi
Joshi, Rishabh
Velury, Sarmishta
Le, Quoc V.
Tan, Qijun
Liu, Yuan
author_facet Ye, Ziyu
Agarwal, Rishabh
Liu, Tianqi
Joshi, Rishabh
Velury, Sarmishta
Le, Quoc V.
Tan, Qijun
Liu, Yuan
contents Current reinforcement learning (RL) frameworks for large language models (LLM) post-training typically assume a fixed prompt distribution, which is sub-optimal and bottlenecks scalability. Prior works have explored prompt evolving, but are often limited to the supervised fine-tuning stage, and prompts are sampled and evolved uniformly without signals. This empirical work presents a paradigm shift: Evolving Alignment via Asymmetric Self-Play (eva), that casts post-training as an infinite game with regret-based signals for 2 players: (i) a creator, who strategically samples and creates new informative prompts and (ii) a solver, who learns to produce preferred responses. eva is the first method that allows language models to adaptively create training prompts in both offline and online RL post-training. The design is simple, easy-to-use yet remarkably effective: eva sets a new SOTA on challenging benchmarks, without any extra human prompts, e.g. it boosts the win-rate of gemma-2-9b-it on Arena-Hard by 51.6% -> 60.1% for DPO and 52.6% -> 62.4% for RLOO, surpassing claude-3-opus and catching up to gemini-1.5-pro, both of which are orders of magnitude larger. Extensive experiments show eva can create effective RL curricula and is robust across ablations. We believe adaptively evolving prompts are key to designing the next-generation RL post-training scheme.
format Preprint
id arxiv_https___arxiv_org_abs_2411_00062
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Scalable Reinforcement Post-Training Beyond Static Human Prompts: Evolving Alignment via Asymmetric Self-Play
Ye, Ziyu
Agarwal, Rishabh
Liu, Tianqi
Joshi, Rishabh
Velury, Sarmishta
Le, Quoc V.
Tan, Qijun
Liu, Yuan
Computation and Language
Artificial Intelligence
Data Analysis, Statistics and Probability
Machine Learning
Current reinforcement learning (RL) frameworks for large language models (LLM) post-training typically assume a fixed prompt distribution, which is sub-optimal and bottlenecks scalability. Prior works have explored prompt evolving, but are often limited to the supervised fine-tuning stage, and prompts are sampled and evolved uniformly without signals. This empirical work presents a paradigm shift: Evolving Alignment via Asymmetric Self-Play (eva), that casts post-training as an infinite game with regret-based signals for 2 players: (i) a creator, who strategically samples and creates new informative prompts and (ii) a solver, who learns to produce preferred responses. eva is the first method that allows language models to adaptively create training prompts in both offline and online RL post-training. The design is simple, easy-to-use yet remarkably effective: eva sets a new SOTA on challenging benchmarks, without any extra human prompts, e.g. it boosts the win-rate of gemma-2-9b-it on Arena-Hard by 51.6% -> 60.1% for DPO and 52.6% -> 62.4% for RLOO, surpassing claude-3-opus and catching up to gemini-1.5-pro, both of which are orders of magnitude larger. Extensive experiments show eva can create effective RL curricula and is robust across ablations. We believe adaptively evolving prompts are key to designing the next-generation RL post-training scheme.
title Scalable Reinforcement Post-Training Beyond Static Human Prompts: Evolving Alignment via Asymmetric Self-Play
topic Computation and Language
Artificial Intelligence
Data Analysis, Statistics and Probability
Machine Learning
url https://arxiv.org/abs/2411.00062