Reproduce “Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents” on OpenResearch

About this paper

Caiqi Zhang, Xiaochen Zhu, Chengzu Li, Yulong Chen, Dharshan Kumaran, Nigel Collier

XConf estimates an LLM’s confidence by retrieving outcome-labeled records of its own previous attempts and combining their empirical success rate with a reflective confidence judgment. Across reasoning, coding, multimodal, and agent benchmarks, it generally improves calibration and selective prediction over verbalized confidence and self-consistency, while requiring no weight updates or token-probability access.

Key insights

• Model-specific experience adds information beyond general task difficulty: conditioning retrieval on stated confidence is central, with ablations showing notable AUROC losses when this signal is removed.
• Historical outcomes help particularly when failures are silent or plausible, as in AppWorld and SWE-bench, while transfer across models is weaker because different models exhibit different failure patterns.
• Confidence estimation can improve with the size of the experience bank, but depends on reliable external grading and faces cold-start, staleness, and retrieval-quality constraints.

Summary by alphaXiv.