Cambridge XConf Beats 10-Sample Self-Consistency on 23 of 24 AUROC

The paper's opening move is a rejection: a model's current inference is not a sufficient basis for judging its own confidence. XConf, described in an arxiv preprint dated September 15, 2026 by six University of Cambridge researchers led by Caiqi Zhang, adds a second signal — the model's own record of graded past attempts. (No em-dash.)

The mechanism runs in two stages. Recall retrieves past episodes on similar tasks met with a similar stated confidence, and reads off their historical success rate. Reflect then shows the model that record, has it name "its recurring failure mode", and asks it to restate a confidence informed by its own track record.

The headline result is efficiency at least as much as accuracy. Across nine benchmarks spanning reasoning, coding, multimodal QA and interactive agents, and four models from three families, XConf "beats or matches ten-sample self-consistency in discrimination (AUROC) on 23 of 24 comparisons, with much lower calibration error (ECE), at a tenth of the generation cost." The estimator is "format-general, requiring no logit access or weight updates, and costs only one answer generation," which matters for closed API models where token probabilities are hidden.

For agent operators, the selective-prediction number is the concrete one. Abstaining on "the 10% least-confident episodes raises the delivered success rate by up to 8.7 points on agent tasks." That is the sort of lever a product team can wire straight into a human-handoff queue.

The abstract prints no per-benchmark AUROC breakdown and does not name the four tested models. It also says nothing about how the episode store gets seeded before Recall has any verified history to draw on.

It lands into a run of hallucinations work our tracker has surfaced this week, alongside the same-day Jevlike option scorer.