Llm agents learn better from critiques plus corrected actions
Beyond Corrected Actions: Training LLM Agents on Textual Critiques
Computation and LanguageArtificial Intelligence
Summary
Large language models (LLMs) that act like agents often make mistakes during multi-step tasks. The authors show that instead of just teaching these models the right action, giving them explanations about why their action was wrong helps them learn better. They created a training method called CAFT that adds these textual critiques alongside corrected actions as feedback. Models trained this way performed about 8% better on several tasks compared to just learning from corrected actions alone. This approach also helps when fine-tuning models before applying reinforcement learning.
What this means in practice
- •For software developers for chatbots: Build chatbot agents that improve decision-making by training on both corrected actions and explanations of errors.
- •For customer service automation teams: Develop more accurate automated agents for complex interactions by integrating textual critiques into agent training.
Authors
Thomas Zeng, Zelei Cheng, Mingyang Zhou, Sambit Sahu, William Campbell, Genta Indra Winata
Abstract
Large language models (LLMs) are increasingly used for multi-turn agentic tasks that require interactions with users and environments over extended sequences of actions. A common approach to improving LLM agents is to use more capable expert models to provide training supervision, either by behavior cloning the expert's trajectories or by using the expert to correct the actions in trajectories generated by the weaker student model. However, expert LLMs can provide richer supervision than actions alone: they can directly generate textual critique that explain why specific actions are incorrect. We study whether training on such feedback benefits LLM agents and propose Critique-Augmented Agentic Fine-Tuning (CAFT), a framework that elicits textual expert feedback on student-generated trajectories. For each incorrect action, the expert yields both a corrected action and a textual critique explaining the student's mistake--which we then use as the training target for fine-tuning. Across four student models from the Nemotron 3 and Gemma 4 families on the agentic domains $τ^2$-Bench Airline+Retail and WebShop, adding critiques to action-level supervision improves performance by 8% on average over action-only distillation. Additionally, CAFT checkpoints can also provide a better warm start for reinforcement learning than expert-trajectory distillation. Our results demonstrate that textual critiques can strengthen supervision for multi-turn LLM agents.