SENTINEL: Failure-Driven Reinforcement Learning for Training Tool-Using Language Model Agents

In three linesSENTINEL is a failure-driven reinforcement learning framework that improves tool-using LLM agents by converting their failures into targeted training tasks. On Tau2-Bench Retail with Qwen3-4B-Thinking-2507, the method increases Pass@1 from 66.4 to 74.9 through a Controller-Proposer-Solver loop that analyzes recurring error patterns.

Summary generated by Claude — human-verified