Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents

Outcome-Grounded Confidence Calibration aligns LLM agent confidence scores with final task success rather than per-step correctness. Using GRPO and proper scoring rules on Qwen2.5-7B, the method calibrates step-wise verbalized confidence against eventual episode outcomes in ALFWorld, TextWorld, and WebShop environments.