The paper proposes XConf, a confidence-estimation method that grounds an LLM’s calibrated confidence in a stored record of its own past graded episodes rather than only the current inference trace. A Recall stage retrieves similar past episodes and their historical success rate, and a Reflect stage has the model restate its confidence informed by that track record, at the cost of a single generation. Across nine benchmarks and four models, XConf matches or beats ten-sample self-consistency in discrimination on 23 of 24 comparisons with much lower calibration error, and abstaining on its least-confident 10% of agent tasks raises delivered success rate by up to 8.7 points.
