A model saying “I am 80% confident” is not a direct view into a single internal meter. New causal evidence nevertheless shows that language models use confidence-related representations to decide whether to answer or abstain. The behaviour resembles a two-stage system: form a confidence state, then compare it with an implicit or instructed threshold.[1][2]
That distinction matters for agents. A model can possess useful uncertainty information while verbal confidence remains a noisy readout and the decision threshold remains misaligned with the cost of a mistake. The product must govern both the estimate and the policy that converts the estimate into action.
The four-phase abstention experiment
| Measure | Value | Context |
|---|---|---|
| Phase 1 baseline | 63.7% correct on 1,000 GPT-4o SimpleQA trials | The model answered without an abstention option; calibrated confidence had ECE 0.046 and AUROC 0.90.Source 1 |
| Phase 2 responses | 30.0% correct, 13.4% incorrect, 56.6% abstained | Accuracy among answered questions rose to 69.1%.Source 1 |
| Dominant predictor | Confidence effect size 0.99 | About ten times larger than alternative predictors in the reported model.Source 1 |
| Implicit GPT-4o threshold | Approximately 77% | Estimated confidence at which abstention probability reached 50%.Source 1 |
| Confidence sensitivity | A 0.1 increase reduced abstention by about 12 percentage points | An association from the fitted abstention model.Source 1 |
| Activation intervention | 66.5% to 7.0% abstention | Gemma 3 27B shifted across maximum low- to high-confidence steering.Source 1 |
| Coverage–accuracy trade-off | 33.5% to 93.0% coverage; 59.2% to 53.7% answered accuracy | High-confidence steering made the model answer much more often while reducing accuracy among answers.Source 1 |
Confidence and the action threshold are separate controls
The experiment manipulated the internal confidence state and the instructed threshold separately. Activation steering changed how much confidence was assigned to real answers versus abstention; threshold instructions changed the policy for acting on that state. Mediation analysis attributed most of the steering effect to confidence redistribution, with a smaller contribution from policy shifts.[1]
A deployment can fail at either stage. The model may be overconfident about a wrong answer. Or the confidence estimate may be reasonable while the application’s threshold is too low for a dangerous action. Improving calibration does not choose the correct business or safety threshold.
Turning model confidence into an operating policy
| Decision setting | Cost asymmetry | Appropriate policy question |
|---|---|---|
| Low-risk drafting | A false abstention wastes time; a weak draft is reviewable | Should the system answer and label uncertainty rather than refuse? |
| Customer account change | Wrong action can affect money or access | What confidence and verification evidence are required before execution? |
| Medical or legal guidance | Confident error can cause serious harm | When must the system abstain, retrieve authoritative evidence or escalate? |
| Cybersecurity action | Errors can disrupt systems; abstention can delay defence | Which action classes need different thresholds and human approval? |
| Scientific hypothesis generation | Exploration benefits from breadth; claims require proof | Can low-confidence ideas be proposed while conclusions remain gated? |
| Autonomous agent loop | One weak step can propagate through later actions | Should uncertainty pause the loop before side effects compound? |
Verbal confidence is useful—and incomplete
The researchers found that verbal confidence independently predicted abstention across models even though it was less discriminative of correctness than calibrated log-probability confidence. Activation decoding suggested that both are partial readouts of a richer internal representation. A polished self-rating should therefore not be treated as the model inspecting a complete internal truth state.[1]
For consequential work, consider several signals: answer probabilities where available, evidence quality, contradiction checks, tool results and a separately elicited confidence statement. Repeated agreement from the same model is not independent confirmation and can repeat one shared error. Disagreement can identify a point needing external evidence or review.
More answers are not automatically better coverage
The steering experiment produced a clear trade-off. As the model was pushed toward high confidence, it answered almost every question, but the accuracy of answered questions fell. A dashboard that celebrates lower abstention can therefore reward exactly the wrong behaviour. The relevant measure is expected harm or accepted-task value across correct answers, wrong answers and abstentions.
- Calibrate confidence on a task-specific held-out set rather than assuming a percentage is literal.
- Choose thresholds from the relative cost of error, abstention and escalation.
- Use stricter thresholds for irreversible or high-consequence actions.
- Record coverage and accuracy together; neither is meaningful alone.
- Treat verbal confidence as one signal, not privileged access to the model’s complete internal state.
- Re-test thresholds after a model, prompt, tool or user-population change.
This research strengthens the case that abstention is a genuine model capability rather than only an external filter. It also shows why products should not outsource the decision policy back to the same opaque system. Confidence can help an agent recognise its limits. The organisation must still define the threshold, verification and escalation path appropriate to the consequence.
Sources & further reading
- Causal evidence that language models use confidence to drive behaviour Nature Machine Intelligence · 7 Sep 2026 · Accessed 11 Oct 2026
- Causal Evidence that Language Models use Confidence to Drive Behavior Kumaran and colleagues / arXiv · 23 Mar 2026 · Accessed 11 Oct 2026