Steering a model's internal confidence moves how often it refuses to answer

Activation steering is where the causal weight sits. Correlational work cannot separate the model gating on its own confidence from confidence and abstention both tracking how hard the question is; an intervention can. The authors pushed the internal confidence signal up and down, and abstention moved inversely, falling when confidence was boosted and rising when it was suppressed ^[3]. Their mediation analysis is the second half of that argument, indicating the behavioural change travelled through a redistribution of confidence rather than around it ^[3].

Phase 2 supplies the shape of the policy. Models behave as though they hold an implicit threshold on internal confidence, and the confidence term outweighs alternative mechanisms by roughly an order of magnitude ^[2], which puts those alternatives near a tenth of the effect size ^[11]. The abstract states that ratio without absolute values ^[12], so the direction of control is better established here than its gain. Earlier work in this area got abstention by bolting a threshold on after the fact or by fine-tuning for it ^[9]; the claim being tested here is that the native signal was already doing the job ^[10].

The awkward part for anyone planning to tune this is which signals are actually reachable. Verbal confidence, elicited in a separate forward pass, independently predicted abstention in every model tested, even though it discriminated correct from incorrect answers less well than the token probabilities did ^[5]. The abstract leaves those models unnamed ^[16]. Activation decoding indicated that both measures are lossy read-outs of a richer internal representation ^[6]. So the dial within reach is a proxy for the quantity doing the work, and part of the gate runs on the poorer indicator of being right. Push a refusal threshold up through that component and you remove answers; it does not follow that you remove the wrong ones.

This result is scoped to four-option multiple choice drawn from a factuality dataset ^[7]. A fixed option set with one correct answer gives confidence something clean to be about, and gives the experimenters a correctness label to score discrimination against. An agent deciding whether to call a tool, or a model deciding whether to answer a clinical question of the kind the paper cites as its motivating case ^[14], has neither. The authors themselves frame the work as mattering because models are becoming autonomous agents that must recognise their own uncertainty ^[8], a statement about why the question is worth asking, separate from evidence that the multiple-choice result carries into that setting.

On what is supplied: abstention on these items is a threshold policy over an internal confidence representation ^[15], the representation can be moved on purpose ^[3], and instructing a model to abstain at a stated confidence level shifts its behaviour in the instructed direction ^[4]. That is enough to treat refusal rate as a quantity with a control input. Specifying a target rate and holding it inside an agent loop is a different question, one this design did not test.