Confidence Comes From Experience: What XConf Changes About How We Measure LLM Confidence

What happened

Last Updated on September 25, 2026 by Editorial Team Author(s): luisacsfreitas Originally published on Towards AI. Confidence Comes From Experience: What XConf Changes About How We Measure LLM Confidence Anyone who has put an LLM in charge of a real decision knows the question that comes right after the demo: when do we trust it? Routing an email to the right team, approving a patch, answering a customer without review. In all of these we need a number that says “this one can go, this one goes to a person”. The trouble is that the number usually comes from the model itself, and the model is not a good judge of itself. On 15 September 2026, a team from the University of Cambridge and Google DeepMind published “Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents” (Zhang, Zhu, Li, Chen, Kumaran and Collier). The method, called XConf, is simple to state and changes the starting point: confidence is not read from the current answer. It is read from the history of past answers. In this post I explain the idea, what the results actually show, how it compares with the rest of the confidence-estimation toolbox, and what it takes to use it in a real system, including the risks the paper does not cover. TL;DR The usual methods (verbalized confidence, token probabilities, self-consistency) only look at the current answer. When a model is wrong the same way every time, all of them say “confident”. XConf keeps a bank of past, graded episodes and asks: on similar tasks, when the model was as sure as it is now, how often was it right? With a single generation, it matches or beats ten-sample self-consistency in 23 of 24 comparisons, with much lower calibration error. It needs no logit access, trains nothing, and works the same for multiple choice, code and agents. The one condition everything rests on: the right/wrong labels must come from outside the model. With labels produced by the model itself, the method gets worse. 1. The problem: confidence only reads the moment There are three families of methods for putting a confidence number on a black-box LLM: Verbalized confidence. Ask the model: “from 0 to 1, how sure are you?”. Cheap and works on any task, but models are systematically overconfident and sensitive to how the question is phrased. Token probabilities (logprobs, P(True)). The probability the model assigned to its own answer. It requires logit access, which most frontier APIs do not give, and it measures uncertainty about the next token rather than about whether the claim is true. Self-consistency. Ask several times with temperature and see whether the answer changes. It was the strongest black-box method available, but it costs N generations per request and does not map cleanly onto code or agent trajectories. All three have one thing in common: they only look at what the model is doing right now. And they share a blind spot. If the model has a stable misconception (it reads “stop paying” and always thinks payments, when in your company that is a cancellation), it gives the same answer every time, with the same certainty and high probabilities. All three methods report “confident” on an error. Legenda: Figure 1. Left: the usual methods only read the current answer. Right: XConf reads the history of similar episodes. 2. The idea: confidence as an observed frequency The paper draws on decades of research on human metacognition. We do not judge our confidence only by re-inspecting the reasoning we just did. We also remember how similar situations turned out. A student who has solved a hundred determinants trusts the result without re-checking. The same student, facing a hard inequality, writes the answer already expecting it to be wrong, because they remember how often such proofs collapsed on the last line. XConf formalises this. Instead of asking the model what it feels, it asks the record what happened. 3. How it works Figure 2. Two readings of the same record: Recall (statistical) and Reflect (verbal). The experience bank. Every time the model solves a task, an episode is stored with five fields: the task; a short reflection by the model on its own solution, written before the outcome is known; the confidence the model stated; the graded outcome (right or wrong), once it arrives; a lesson, written by the model after it learns the outcome. The lesson is quarantined: it is never shown to the model while it assesses a solution that has not been graded yet. The authors are explicit about why: knowing the outcome biases self-judgement in ways that instructions alone do not fix. Recall: a confidence-conditioned hit rate. For a new task, XConf retrieves the most similar episodes from the bank. The important detail is that the search key has two parts: the task embedding and the stated confidence. It does not look for “similar tasks”, it looks for “similar tasks where the model felt equally sure”. It takes the 50 nearest neighbours and computes the fraction in which the model was right. The paper shows this conditioning is not decoration: removing stated confidence from the key costs 0.08 AUROC on reasoning and 0.12 on agents. There is one more subtle detail: similarity is not raw embedding cosine. It is measured in a space rescaled using the bank’s own graded outcomes, so that “similar” means “fails for the same reasons” rather than just “is about the same topic”. Reflect: the model reads its own track record. The retrieved neighbours are shown to the model as short cards (task, stated confidence, outcome, lesson). The model first has to name the recurring failure mode it sees, and only then restates a confidence. It is a short call that does not re-solve the task. Combine. The final confidence is the plain average of the two readings: ½ × (Recall + Reflect). Two properties make this practical. When the bank is sparse around a task, the neighbours are only weakly similar and Recall […]

What to do next

Compare the hosted model pages first, then check the related tools and buyer guides before changing workflow standards.