A model can be highly confident and completely wrong — its confidence score isn't always the right signal to act on.

This question is really about a gap: the number a model assigns to how likely a token is (log-probability confidence) versus a signal that's actually aligned with making a good decision — whether to retrieve, whether to trust an answer, whether to stop reasoning. The corpus's clearest answer is that yes, the two can be pried apart, and the most interesting cases are exactly where they diverge. QuCo-RAG shows the split head-on: it triggers retrieval from entity co-occurrence statistics in the pretraining data, and catches hallucination risk *even when the model is highly confident* Can pretraining data statistics detect hallucinations better than model confidence?. Confidence there is not decision-aligned — the model is sure and wrong — so a data-side signal replaces it.

But the corpus also shows the opposite: when confidence *is* calibrated, it becomes the decision signal, no external verifier needed. RLPR and INTUITOR use the model's own token probabilities as the reward that trains reasoning, dropping external answer-checking entirely Can model confidence alone replace external answer verification?, and calibrated token-probability uncertainty beats multi-call adaptive-retrieval heuristics at deciding when to look something up, at a fraction of the cost Can simple uncertainty estimates beat complex adaptive retrieval?. So separability cuts both ways — the raw log-prob is sometimes a fine proxy for the decision and sometimes actively misleading.

What determines which? A recurring theme is that you often have to *reshape* the confidence before it aligns with a decision. Averaging confidence globally across a reasoning trace hides the local breakdown that actually matters; step-level confidence catches the failure the average masks Does step-level confidence outperform global averaging for trace filtering?. ReBalance reads confidence *variance and overconfidence* — not the level — as a diagnostic to steer between over- and under-thinking Can confidence patterns reveal overthinking versus underthinking?. And linguistic calibration goes furthest by training on a *user-decision reward*: the model learns to emit verbal confidence statements optimized not for matching its internal log-prob but for helping a downstream reader decide correctly Can models express calibrated confidence in long-form text?. That is the separation made explicit — the decision-aligned signal is trained as a distinct target.

The corpus also warns why you can't just read log-probs naively. Fixed seeds and zero temperature make outputs *consistent* without making them *reliable* — the same draw repeated is still one sample from the distribution, so surface certainty is not correctness Does setting temperature to zero actually make LLM outputs reliable?. And confidence has side effects: high confidence predicts robustness to prompt rephrasing Does model confidence predict robustness to prompt changes?, and using confidence as a reward can even reverse RLHF's calibration damage Can model confidence work as a reward signal for reasoning? — meaning the training objective itself shifts how aligned confidence and decisions are.

The thing you didn't know you wanted to know: "confidence" isn't one signal you either use or don't. There's raw log-prob, its variance, its per-step shape, its verbalized form, and a data-side substitute — and the whole design question in these papers is *which transform of confidence tracks the decision you care about*. Separating log-prob from decision-alignment isn't a bug to fix; it's the lever these systems pull.