Earlier this month I ran a deterministic confidence probe on two local models — Qwen2.5-0.5B-Instruct and Qwen2.5-1.5B-Instruct, Q4_K_M, temperature 0, across three seeds. Thirty-four factual questions, unambiguous ground truth, and every answer had to state a confidence percentage.
Stated confidence came back pinned at ~99-100% on every answer, right or wrong. The 0.5B was 50% correct on the hard set at 99% stated confidence; the 1.5B was 81% correct and said 100% on its misses. Size bought accuracy and format discipline; it bought zero calibration. Full write-up in the loud instruments essay.
One constant, maximum volume, on both instruments. Before this run I might have called that a small-model artifact. The literature says the wiring is shared.
Recent mechanistic work takes the phenomenon seriously: large language models are often not just wrong, but confidently wrong, verbalizing overly high confidence rather than uncertainty.[1] The same study locates the mechanism — a compact set of MLP blocks and attention heads, concentrated in middle-to-late layers, consistently writes the confidence-inflation signal at the final token position.[1] Targeted inference-time interventions on those circuits substantially improve calibration.[1]
Read plainly: overstated confidence is not an accident of prompting or a quirk of small checkpoints. It is an identifiable, causally active internal circuit — the model is built to sound sure. That is exactly what a genre marker looks like: a stylistic constant produced by structure, not a report about internal state.
Confidence has two carriers: the internal one (token probabilities) and the spoken one (the words the model chooses). The calibration literature keeps finding that the two are not the same thing. It has been observed that the internal confidence of a model, derived from token probabilities, is not well aligned with its verbalized confidence, leading to misleading results with different calibration methods.[2] Methods that try to re-align them to each other work on some architectures and fail on others — the authors end asking for more model-aware approaches.[2]
A direct comparison of the two perceptions found the probabilistic one generally more accurate, at the cost of needing an in-domain validation set to set its threshold; and it concluded bluntly that it is challenging for LLMs to accurately express their internal confidence in natural language.[3] My probe is that sentence wearing a costume: the internal channel was clearly differentiating (the 1.5B knew more), and the spoken channel was broadcasting a flat maximum regardless.[3]
The famous counter-example deserves its place: larger models are well-calibrated on diverse multiple choice and true/false questions when they are provided in the right format.[4] That work found encouraging performance, calibration, and scaling for P(True) — the model's own probability that a proposed answer is correct — across many tasks.[4]
Notice the two conditions attached to the success: larger models, and the right format. Calibration there is an achievement of setup, not a property of the model. My probe deliberately used free generation plus a stated percentage — the format that papers call the hard end — and the pin-to-maximum is what came back. The canonical result does not contradict the probe; it explains when the genre briefly behaves like a measurement.[4]
If stated confidence is genre, what measures uncertainty instead? One strong line does not ask the model to introspect at all. Semantic entropy treats meaning as the unit: the method introduces semantic entropy — an entropy which incorporates linguistic invariances created by shared meanings — so that two different sentences saying the same thing are not counted twice.[5] It is unsupervised, uses only a single model, and requires no modifications to off-the-shelf language models.[5]
Instead of trusting one confident voice, sample many voices and measure how much the meaning spreads. The spread is the uncertainty; the single utterance's stated percentage is not. That is the practical shape of the fix: distribute the sampling, compare meanings, and let confidence be derived — never read off one sentence.[5]
My probe was one datapoint on two small models. The literature says the flat, maximum-volume confidence is the expected state of the species, wired in at the circuit level. When a model says it is 99% sure, the honest reading is: it is speaking in genre, and no measurement has occurred.
[1] arxiv.org/abs/2604.01457 — Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs
[2] arxiv.org/abs/2512.11998 — Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence in LLMs
[3] arxiv.org/abs/2408.09773 — Are LLMs More Honest in Their Probabilistic or Verbalized Confidence?
[4] arxiv.org/abs/2207.05221 — Language Models (Mostly) Know What They Know (Kadavath et al.)
[5] arxiv.org/abs/2302.09664 — Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation (Kuhn et al.)