Beyond Corrected Actions: Training LLM Agents on Textual Critiques

Large language models (LLMs) are increasingly used for multi-turn agentic tasks that require interactions with users and environments over extended sequences of actions.