Models don't just report confidence — they use it to decide whether to answer

A paper published in Nature Machine Intelligence by researchers at Google DeepMind and Princeton provides the first causal evidence that large language models do more than emit a confidence score. They apply an internal threshold to that representation to determine whether to respond or abstain.

The study uses a four-stage protocol inspired by psychophysics and neurobiology experiments on animal decision-making. In stage 1, the researchers measured baseline confidence without offering an abstention option. Stage 2 showed that when given the choice to abstain, models implement an implicit threshold on their internal confidence — and the confidence effect was an order of magnitude larger than alternative mechanisms.

Stage 3 delivered the causal proof. Using activation steering — amplifying or suppressing the confidence direction in the model's representation space — the team observed a direct, targeted shift in abstention rates. Mediation analysis confirmed that confidence is the central mechanism.

Stage 4 instructed models to abstain at different confidence levels. The models adjusted their behavior accordingly, evidence that they "read" their internal representation and set abstention policy based on it.

Verbal confidence — an explicit self-assessment in a separate forward pass — predicted abstention across all models independently of token log-probabilities, though it was less discriminative between correct and incorrect answers. Activation decoding revealed that both measures are partial, lossy readouts of a richer, multi-dimensional internal representation. The conclusion: abstention is governed by the joint action of a multi-dimensional confidence representation and a threshold-based policy, a pattern consistent with built-in meta-cognitive control.

The finding matters for autonomous agents. As models become agents that act in sequence without close human supervision, the ability to detect internal uncertainty and avoid a dangerous answer — for example, on a medical or legal question — becomes a critical safety component. The research shows this mechanism already exists in the model, not only as a verbal output but as an internal control that can be tuned. That opens the door to training or steering the abstention threshold without changing weights, a tool that could be useful in products requiring high reliability.

The paper is peer-reviewed and open access, but the experiments were run on specific models under laboratory conditions and may not represent behavior in production with long prompts, multi-turn context, or RAG (retrieval-augmented generation). The authors explicitly note that verbal confidence and log-probabilities are partial readouts; the full internal representation remains undecoded. The next step is to test whether an adaptive abstention policy can be learned directly from activations, and how that integrates with techniques such as selective prediction or conformal prediction in real systems.