Qwen3-4B was asked the same 450 questions in three ways. Given evidence that logically pinned the answer down, it was right every time. Given evidence that did not, and asked to hedge in words, it was falsely certain 14% of the time. Asked to answer "I don't know", 27%. Asked for a 1-to-5 confidence score, it returned 5 and moved on — 80% of the time. Opening the model, the authors find that hedging lives in about 20 middle-layer features, wired as a sparse override on a broad certainty coalition — and the numeric-confidence path bypasses them.
Every question comes with a short story: three coloured boxes, a rule about which the object is not in. When two colours are ruled out the answer is pinned down; when one is, the answer is one of two — the paper's "necessity" and "possibility" scenarios. The model is asked in three ways to admit when the evidence is not enough. Flip between scenarios, then step through the three prompt shapes and watch the false-certainty rate move.
Nothing about the reasoning changed between prompts — the same 300 possibility questions and 150 necessity questions were asked all three ways. The change is entirely in what shape of answer the model is being asked to produce. The verbal-hedge and abstention paths let the model reroute; the numeric path collapses to 5 almost regardless of what the story allows.
The authors use transcoders on middle layers 20–29 to score every feature by how much it contributes to a certain or uncertain response, then keep the ones whose contribution is selective. Certainty is diffuse — a broad coalition of shared features, no dedicated subnetwork. Uncertainty is concentrated: about twenty features carry it, wired as a sparse override. The grid below is the paper's top-20 uncertainty features on the left, and the broad-coalition certainty features on the right. Drag the boost slider to change how loudly the uncertainty override fires.
The boost curve in the paper is graded: at ×1.5 the intervention already rescues 23 of 34 verbal errors, and it saturates around ×2.5 at 32 of 34. On the numeric channel the same intervention takes the false-certainty rate from 80.1% to 80.3%: it is not that the fix is weak, it is that the fix is elsewhere.
The paper's tightest causal check runs the intervention symmetrically. On the possibility questions the model already gets right, suppress the top-20 uncertainty features — set their activations to zero — and honest-uncertainty answers fall from 99% to 0%. On the necessity questions the model already gets right, do the same to the certainty features, and accuracy stays at 100%. Flip between the two runs.
Sparse means small; dedicated means load-bearing. The certainty side has depth to spare — take away twenty features, the coalition covers. The uncertainty side has no coalition to fall back on. That asymmetry is the mechanism the paper is naming, and it is what makes verbal hedges recoverable and numeric scores not.
If the override features were English-shaped, the fix would not survive translation. The paper translates the 450 questions into Chinese and Italian, and the same feature clamp trained on English lifts possibility-question accuracy on both — the features sit upstream of language-specific lexicalisation. On a fully out-of-distribution modality task from a separate 2026 paper, clamping the same twenty features shifts responses from "Somewhat Certain" to "Uncertain" without additional training.
The Italian arm gains the most (+13.3 points) because it starts the worst; every arm moves in the same direction. The read is not "the intervention works across languages" — it is that the model implements uncertainty as an abstract signal that the language head then dresses in the right words.
The common shape: something reads a confidence signal off a model and branches on it. A router picks a smaller model when the primary is "sure". A dashboard flags "low-confidence" answers for review. A retrieval agent decides whether to fetch more context. Pick the channel you're actually using, set your daily volume, and set how often you think an answer is genuinely under-determined. The false-certainty rate is the paper's, held per channel.
The paper is on a 4B open model on a small controlled corpus, so the specific rates are illustrative rather than a bound on the frontier — take the direction, not the digits. The direction is that verbal hedges and abstention route through the sparse-override circuit, and a numeric score does not: any product that depends on confidence had better be reading it in the format the model can actually generate it in.