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

Why the current inference is not enough

Every existing confidence estimator reads only the current inference: it introspects on the answer just produced, scores its token probabilities, or resamples it and counts agreement. Each estimate is made as if the model had never faced a similar task before. People do it differently: a student who has just solved yet another determinant exercise trusts the result without rechecking a step, because problems of this kind have never let them down. XConf gives a language model the same resource: its own track record.

The final estimate is the mean of the two readings, one statistical and one verbal, with equal weight.

Grounded in how humans calibrate

Five decades of metacognition research paint one picture, and each finding is a design choice here:

Every family before it reads the current inference

Verbalized confidence introspects on the answer; likelihood methods score its token probabilities; consistency resamples the same task ten times and counts votes; post-hoc calibrators remap one of those scores. None of them consults the model’s own history at inference time. XConf estimates confidence from that history: the model’s accumulated experience.

Results

The same recipe across reasoning, code, multimodal QA, and interactive agents (ScienceWorld, AppWorld, SWE-bench Verified), on Gemini 2.5/3.5 Flash, Claude Sonnet 4.6, and Qwen3.5‑397B. The margin is largest exactly where self-consistency does not naturally apply: code and agents. On AppWorld, where a failed rollout looks clean, verbalized confidence averages .72 across the four models while XConf reaches .86.

Explore the full matrix

Every method the paper compares, on every model column. Pick a dataset; within each group, rows sort by mean AUROC. Bold marks the best AUROC in a column and anything within .01 of it, the paper’s on-par band. ECE is not bolded: labeled calibrators that flatten scores reach a degenerate near-zero ECE.

Experience travels, and mixing banks is safe

Banks for different benchmarks can be merged without curation and without penalty: pouring every other dataset’s bank into the target’s changes AUROC by at most .01. A bank of the same kind transfers outright. A bank built from a different coding corpus (R2E-Gym) calibrates SWE-bench Verified as well as SWE-bench’s own bank, and merging the two beats either bank alone.

Borrowing has a measurable cost in two cases. Another model’s bank on the very same tasks loses .03 to .06 AUROC: what a record shares is which problems are hard, and what it cannot share is a model’s own failure pattern. This is the self versus other asymmetry found in human metacognition, and the reason XConf defaults to the actor’s own record. Across task families the median cost is .08, and the worst target is code, for the plain reason that non-code banks contain no code experience.

The cost of borrowing a bank. AUROC lost relative to the target’s own bank, retrieving from another dataset’s bank (left) or another model’s bank on the same tasks (right).

More experience, better calibration

Holding everything else fixed and growing the bank from a hundred episodes to the full pool traces a rising curve on all three agent domains; on AppWorld with Claude Sonnet 4.6, AUROC climbs from .628 to .810. Where a curve flattens, it flattens because the bank is already dense: MMLU-Pro levels off at twelve thousand episodes, while BBEH and the agent domains, where episodes are fewest, are still climbing at the full pool. More deployment therefore keeps buying better calibration exactly where experience is scarce today.

AUROC as the bank grows (log scale), all four models: two agent domains, plus BBEH and MMLU-Pro under the same protocol.

Selective prediction

Rank episodes by estimated confidence and act on the score: ship the confident set, hand off the tail. Abstaining on the least-confident 10% raises delivered accuracy by 4.8 points, averaged over all thirty-six model and dataset cells, and every cell gains. The largest gain is 8.7 points on AppWorld with Gemini 3.5 Flash, from .805 to .892. Drag the slider to see what each estimator delivers.

What else holds

It estimates more than question difficulty. Hold each question’s difficulty fixed, by how the other three models did on it, and XConf still separates correct from incorrect attempts: stratified AUROC .79, above chance in 35 of 36 cells, and still .05 ahead of verbalized confidence.

It survives the actor shrinking. On Qwen3.5 at 397B, 27B, 9B and 4B, task accuracy falls steeply while the estimate barely moves. The statistical Recall reading is nearly size-invariant; only the verbal Reflect reading degrades below 27B. The split has a plain reason: Recall counts outcomes in the record and asks nothing of the model’s self-knowledge, whereas Reflect asks the model to read its own record, and smaller models are markedly weaker at assessing their own outputs.

It tolerates noisy labels, but not the model’s own. A weak but independent grader, an LLM judge that agrees with gold labels .91 of the time, keeps most of the value: the judge-labeled bank still ties ten-sample self-consistency at a fifth of the cost. A bank labeled by the model itself scores below a bank with no outcome labels at all.

BibTeX

@article{zhang2026xconf,
  title  = {Confidence Comes from Experience: Experiential Confidence
            Estimation from Reasoning to Agents},
  author = {Zhang, Caiqi and Zhu, Xiaochen and Li, Chengzu and Chen, Yulong
            and Kumaran, Dharshan and Collier, Nigel},
  journal = {arXiv preprint arXiv:2609.17708},
  year   = {2026}
}