Latest AI Paper

Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents

XConf estimates a model's confidence by recalling its own graded past episodes instead of only inspecting the current inference. It retrieves similar past tasks with similar stated confidence, reads off historical success rates, and has the model reflect on recurring failure modes. Across nine benchmarks and four model families, XConf beats or matches ten-sample self-consistency on 23 of 24 AUROC comparisons with much lower calibration error at a tenth of the generation cost.

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