Language Models Can Learn from Verbal Feedback Without Scalar Rewards

2025-09-26

AI Summary

Existing LLM feedback learning relies on scalar rewards, leading to loss of semantic richness in natural language feedback and reward-scale imbalance. This paper proposes Feedback-Conditioned Policy (FCP), the first approach to treat natural language feedback directly as a generation condition—not a reward signal—thereby reformulating feedback learning as a conditional language modeling task. Methodologically, FCP integrates offline supervised fine-tuning (on response-feedback pairs), online feedback sampling, and policy self-bootstrapping, entirely eliminating the need for scalar reward design. Experiments demonstrate that FCP significantly enhances model understanding of and responsiveness to complex, fine-grained feedback, enabling sustained improvement in generation quality across multi-turn interactions. The implementation is publicly available.

Technology Category

Natural Language Processing: GenerationMachine Learning: Large Multimodal Models (LMMs)Search and Optimization: Learning to Search

Application Category

User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendationSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

LLMs are often trained with RL from human or AI feedback, yet such methods typically compress nuanced feedback into scalar rewards, discarding much of their richness and inducing scale imbalance. We propose treating verbal feedback as a conditioning signal. Inspired by language priors in text-to-image generation, which enable novel outputs from unseen prompts, we introduce the feedback-conditional policy (FCP). FCP learns directly from response-feedback pairs, approximating the feedback-conditional posterior through maximum likelihood training on offline data. We further develop an online bootstrapping stage where the policy generates under positive conditions and receives fresh feedback to refine itself. This reframes feedback-driven learning as conditional generation rather than reward optimization, offering a more expressive way for LLMs to directly learn from verbal feedback. Our code is available at https://github.com/sail-sg/feedback-conditional-policy.

Problem

• Learning from verbal feedback without scalar reward compression
• Treating feedback as conditioning signal for language models
• Reframing feedback learning as conditional generation problem

Innovation

• Treats verbal feedback as conditioning signal
• Uses feedback-conditional policy with maximum likelihood
• Online bootstrapping refines policy with fresh feedback