RLVF: Learning from Verbal Feedback without Overgeneralization

TL;DR

This paper introduces C3PO, a method that incorporates high-level verbal feedback into large language models to prevent overgeneralization, ensuring feedback is applied only in relevant contexts while maintaining original behaviors.

Contribution

C3PO is a novel approach that uses synthetic preference datasets and constrained fine-tuning to accurately apply verbal feedback without overgeneralizing to irrelevant scenarios.

Findings

C3PO reduces overgeneralization by 30%.

It effectively applies verbal feedback in relevant contexts.

Performance is comparable to in-context baselines.

Abstract

The diversity of contexts in which large language models (LLMs) are deployed requires the ability to modify or customize default model behaviors to incorporate nuanced requirements and preferences. A convenient interface to specify such model adjustments is high-level verbal feedback, such as "Don't use emojis when drafting emails to my boss." However, while writing high-level feedback is far simpler than collecting annotations for reinforcement learning from human feedback (RLHF), we find that simply prompting a model with such feedback leads to overgeneralization of the feedback to contexts where it is not relevant. We study the problem of incorporating verbal feedback without such overgeneralization, inspiring a new method Contextualized Critiques with Constrained Preference Optimization (C3PO). C3PO uses a piece of high-level feedback to generate a small synthetic preference dataset specifying how the feedback should (and should not) be applied. It then fine-tunes the model in accordance with the synthetic preference data while minimizing the divergence from the original model for prompts where the feedback does not apply. Our experimental results indicate that our approach effectively applies verbal feedback to relevant scenarios while preserving existing behaviors for other contexts. For both human- and GPT-4-generated high-level feedback, C3PO effectively adheres to the given feedback comparably to in-context baselines while reducing overgeneralization by 30%.