/p/2026-08-29 · explainer
Paper explainer · 2608.26121 · Asaria, Salomone & Gandhi

It already knows when it is guessing.

Teaching a model to say "I'm not sure" normally starts with a labelling job: ask it thousands of questions, mark which answers were wrong, then fine-tune it to decline exactly those. This paper deletes the labelling job. Rank the training questions by the average log-probability the frozen model gave its own answer, retarget the least-confident slice to a fixed refusal, keep the model's own answer everywhere else, and train a small adapter. Across six open models from 1B to 8B it lands within noise of the labelled version — 0.95 to 1.10 times its hallucination rate at matched coverage. The blind spot is the obvious one: the confident mistakes never trip the wire.

01 · The signal

The probability dips on the facts it gets wrong

A language model states a false fact as fluently as a true one, but the fluency is not uniform underneath. Average the log-probability the model assigned to each token of its own answer and you get one number per question — a doubt score, computed once, from the frozen base model, with no labels and no second model involved. Rank a question set by that number and the wrong answers pile up at the bottom. How reliably is the question this section answers: pick a model and see how well its own confidence separates its mistakes from its hits.

Interactive · pick a modelseparation quality, 50 = coin flip, 100 = perfect
5075100
separation between right and wrong answers
wrong answers before any training

This is a useful signal, not a great one. Every model clears chance comfortably and none is near perfect — the range across the six is 72.5 to 85.8 on a scale where 50 is a coin flip. Note also that it does not simply improve with size: the largest Qwen model separates its own errors slightly worse than the smallest one does.

02 · The recipe

Three ways to build the same fine-tune

The training recipe is identical in all three arms; only the rule choosing which questions get retargeted changes. That is what makes the comparison clean — the difference between arms is a supervision cost, not an architecture. Step through them and read what each does to the training set.

Interactive · step through the three armssame LoRA recipe, same data, different rule for who abstains

      
what it costs to build
what it does to hallucination

Interactive · move the abstention thresholdthe coverage knob illustrative · scores shown are a worked example

The threshold is the only knob, and it is a coverage knob: push it up and the model answers less often but is right more often when it does. Nothing in the pipeline needs to know which answers were actually correct.

03 · The parity

Free supervision, paid supervision, same result

Here is the claim the paper is built to support. On every one of the six models, the label-free arm and the label-supervised arm land within each other's confidence intervals at matched coverage — a ratio of 0.95 to 1.10 between their hallucination rates, with no difference detected anywhere. Below, each model's journey from answering everything to answering selectively.

Interactive · pick a modelshare of given answers that are wrong · before training → label-free arm
0%50%100%
reduction in wrong answers
questions it still answers

One control kills the obvious alternative explanation. Take the same low-confidence questions and, instead of retargeting them to a refusal, train harder on getting them right — the "just memorise more" arm. It stays at the base hallucination rate with near-full coverage. The gain comes from knowing when to stop, not from learning more facts.

04 · The blind spot

What a single-pass confidence score cannot see

Now the part to design around. A doubt score built from the model's own probabilities can only catch the mistakes the model is uneasy about. It is silent on the ones it is sure of — and those are the majority of what is left after training. Toggle between the two halves of the residual.

Interactive · look at the residual two waysshare of all questions after training · what remains wrong
0%20%40%

There is a shape to where the errors live, and it is the one you would guess: rarity. Split questions by how popular the entity being asked about is, and the least popular quartile hallucinates two to three times as often as the most popular. That is the part of your traffic where a confidence gate matters most and where it also has the most to be uncertain about.

Chart · hallucination by how obscure the subject isthe 8B Llama model, best and worst popularity quartiles
0%50%100%

Same model, same prompt format, same day: 32% wrong on well-known entities and 89% wrong on obscure ones. Any single accuracy number you have measured is really a statement about the popularity mix of the questions you sampled.

05 · Your product

What a doubt gate does to your support queue illustrative

Put this in front of a product surface — a knowledge-base assistant, a support deflection bot, any place a model answers a factual question in front of a customer. Set your volume and pick a model tier; the outcome shares are the paper's measured numbers for that model, and the traffic is yours. The trade is not subtle: far fewer wrong answers, far fewer answers.

Interactive · scale it to your trafficcoverage and error rates measured · question volume is yours
wrong answers shown, before
wrong answers shown, after
questions handed to a fallback
correct answers you gave up

The abstained questions are not lost, they are routed: to retrieval, to a larger model, to a search box, to a person. That is what makes the coverage cost payable — and it is why the useful version of this is a gate in front of a fallback, not a model that shrugs at the customer.

The numbers

What the paper actually reports

What this does not show

In practice