2024 TrainingLanguageModelstoSelfCor

From GM-RKB
Jump to navigation Jump to search

Subject Headings: Intrinsic Self-Correction, Self-Correction via Reinforcement Learning, Multi-Turn Reinforcement Learning, Reward Shaping, On-Policy Learning, Distribution Shift, Policy Initialization.

Notes

Thank you for the guidance. I'll revise the bullet points with more appropriate and unambiguous wiki links:

Cited By

Quotes

Abstract

Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Existing approaches for training self-correction either require multiple models or rely on a more capable model or other forms of supervision. To this end, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.

References

;

 AuthorvolumeDate ValuetitletypejournaltitleUrldoinoteyear
2024 TrainingLanguageModelstoSelfCorCosmin Paduraru
Shariq Iqbal
Aleksandra Faust
Feryal Behbahani
Rishabh Agarwal
Avi Singh
George Tucker
Kate Baumli
Disha Shrivastava
Colton Bishop
Aviral Kumar
Vincent Zhuang
Yi Su
John D Co-Reyes
Rebecca Roelofs
Lei M Zhang
Kay McKinney
Doina Precup
Training Language Models to Self-Correct via Reinforcement Learning2024