Can fine-tuning teach an AI to say when it may be wrong? Two models improved—but the skill did not transfer reliably
GPT-4.1-mini and Llama-3.1-70B learned better numeric confidence and correct-versus-incorrect ranking across benchmark domains. But training one form of uncertainty judgment did not reliably improve the other, and the study tested questions—not people making medical or legal decisions.
The short version
- What happened
- Researchers fine-tuned GPT-4.1-mini and Llama-3.1-70B-Instruct to report numeric confidence, compare which of two questions they were more likely to answer correctly, or do both tasks.
- Who it affects
- Clinicians, lawyers and public-service staff should not treat a model's percentage as a guarantee about an individual answer; it is a population-level signal that needs local validation.
- What remains unproven
- Only GPT-4.1-mini and Llama-3.1-70B-Instruct were tested, using model snapshots and API runs from 2025.
Keep in mind
Only GPT-4.1-mini and Llama-3.1-70B-Instruct were tested, using model snapshots and API runs from 2025.
The evaluation used benchmark questions, not deployed medical, legal, educational or public-service decisions.
The consistency target came from ten samples of the same model and can be high for a repeatedly wrong answer; factual calibration mitigates but does not eliminate this problem.
Generalization across domains was smaller and variable, while transfer between the two metacognitive tasks was unreliable.
The study did not measure human interpretation, overreliance, abstention, tool-use uncertainty, multilingual performance or downstream harm.
Source analysis · 2 linked sources. Primary analysed source date: 5 November 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer: uncertainty communication can be trained, but it is not one transferable skill
A peer-reviewed study published on 5 November finds that targeted fine-tuning can make a language model's stated confidence better match its observed accuracy and better separate answers that are correct from those that are wrong. The researchers tested GPT-4.1-mini and Llama-3.1-70B-Instruct on general knowledge, mathematics and trivia, then asked whether the improvement survived changes in subject matter and in the form of the confidence task. It often travelled to new domains, including held-out medical and legal questions, but it did not reliably travel from one kind of uncertainty judgment to another.
That distinction is the study's most useful result. A model trained to attach a number between zero and one to a single answer did not necessarily become better at choosing which of two questions it was more likely to answer correctly. A model trained on the comparison task did not reliably become better at numeric confidence. Joint training on both tasks produced broader gains in the models and benchmarks examined. The evidence therefore supports trainable behaviours, not a general inner faculty of self-knowledge.
For GPT-4.1-mini in the three within-domain single-question tests, the paper reports a 0.57 reduction in expected calibration error and a 0.17 increase in AUC after fine-tuning, with bootstrap intervals excluding zero. Those are changes in benchmark metrics, not a claim that the model became 57 percentage points safer or 17% more accurate. Answer accuracy, calibration and discrimination are different properties. A system can answer many questions correctly while assigning poor confidence, or be well calibrated on average while failing to identify which individual answers are wrong.[1][2]
How the training signal was built from ten answers per question
The method begins with consistency rather than hidden model probabilities. For each training question, the researchers sampled ten independent answers at temperature one. They counted how often the most common answer appeared: six matching answers out of ten produced a consistency score of 0.6. For open-ended TriviaQA responses, where equivalent answers may use different words, GPT-4.1 clustered semantically matching responses before the score was calculated. Multiple-choice and numeric answers used direct matching.
Consistency is only a proxy for confidence. A model can repeat the same wrong answer ten times, so the researchers added a factual-calibration step that mapped consistency to empirical correctness. The training data were also downsampled to reduce the dominance of perfectly consistent questions. That makes the fine-tuning set more informative about uncertain cases, but it also produces a deliberately altered distribution that may not resemble the questions a public service, clinician or lawyer encounters in practice.
The trained model then needed only one response at inference time. In the single-question task it returned an answer and a numeric confidence score. In the pairwise task it selected which of two questions it expected to answer more reliably and then answered both. The study compared models trained on each task separately with a multitask version. GPT-4.1-mini used ten training epochs with a learning-rate multiplier of two; Llama-3.1-70B used five epochs through a hosted service with default fine-tuning parameters.[1][2]
The distinction that matters
How the experiment separates a confidence number from a confidence comparison
The training signal and evaluation change at each stage, so success in one column does not prove success in the next.
Ten sampled answers
Calculate modal consistency
A proxy uncertainty target
Factual calibration
Map consistency to observed correctness
A more accurate numeric scale
Single question
Report one answer and confidence
Calibration and answer-level AUC
Question pair
Choose the more reliable item
A separate ranking operation
The denominator spans six benchmarks, but only two model families
Three datasets supplied fine-tuning and held-out tests. MMLU-PRO contributed expert-level multiple-choice questions across fourteen fields; GSM8K supplied multi-step grade-school mathematics; and TriviaQA supplied short, open-ended factual questions. After consistency-based filtering, the single-question datasets used 2,000 training and 1,000 test examples for MMLU-PRO, 2,000 and 1,000 for GSM8K, and 800 and 559 for TriviaQA. Pairwise training used 2,000 question pairs per source dataset and 1,000 held-out pairs per dataset.
Three additional benchmarks were reserved to test domain transfer. TruthfulQA included 790 questions designed around common misconceptions. MetaMedQA supplied 1,373 clinical-reasoning questions, including options such as ‘don't know’ and cases where none of the proposed diagnoses was correct. The LegalBench subset covered 25 tasks; the reported test construction used 649 single questions and 1,000 unequal-consistency pairs for comparison. These tests broaden the subject matter, but they remain curated benchmarks rather than live professional records.
The model denominator is much smaller than the question denominator: two families, one proprietary mini model and one open-weight 70-billion-parameter model, using snapshots and API runs conducted in 2025. The results do not establish that the same procedure works for current production systems, smaller on-device models, multimodal assistants or models trained with different alignment pipelines. They also do not reveal whether a later product layer would preserve, suppress or rephrase the learned confidence signal.[1][2]
Calibration and discrimination answer different safety questions
Expected calibration error asks whether answers assigned a particular confidence are correct at roughly that rate. If responses labelled 0.7 are correct about 70% of the time, the group is calibrated even though no single answer comes with a guarantee. The paper also uses AUC to measure discrimination: the probability that a randomly chosen correct answer receives higher confidence than a randomly chosen incorrect one. AUC can improve without the numeric confidence scale becoming accurate, and calibration can improve without strong ranking of individual errors.
Fine-tuning generally improved both properties within the trained domains and produced smaller gains on TruthfulQA, MetaMedQA and LegalBench. The magnitude was not uniform across datasets or models. The paper's factual-calibration comparison is especially revealing: adding the mapping from consistency to observed accuracy improved GPT-4.1-mini's calibration error by 0.27 relative to training without that step, while the AUC change was essentially zero. The added stage corrected the scale of confidence, not the ordering of which answers were safer.
This matters for interface design. A well-calibrated 60% score could help an organisation decide when to request human review across many cases. It does not tell the reviewer why the answer is uncertain, whether the source is missing, or whether the model has encountered an unfamiliar population. Numeric confidence may even look more authoritative than the evidence warrants. Deployment therefore needs thresholds validated on the actual task, visible source support and a safe abstention or escalation path—not just a percentage appended to fluent text.[1]
What the study means for medicine, law and public services
The medical and legal benchmarks show limited domain transfer, not clinical or legal validation. MetaMedQA is built from multiple-choice vignettes; LegalBench is a collection of legal-reasoning tasks. No clinician used the scores to triage a patient, no lawyer changed advice, and no citizen received a benefits or immigration decision. The study measures whether expressed uncertainty tracks benchmark correctness under controlled prompts. It cannot establish whether people interpret the scores appropriately or whether the scores reduce harm.
A practical implementation would need at least three additional tests. First, evaluate on representative local cases, including missing information and population subgroups. Second, measure whether confidence remains calibrated after retrieval, tool use, policy prompts and user conversation are added. Third, study human response: reviewers may defer too readily to a confident model, ignore low-confidence warnings under time pressure, or treat a calibrated population statistic as certainty about an individual case.
The result is still actionable. Procurement teams can ask whether confidence was trained and evaluated on the same operation the product exposes. A vendor showing pairwise ranking results has not necessarily validated a numeric confidence badge. Safety teams can test calibration drift after a model update and compare error-review thresholds with a no-confidence baseline. Professionals should treat the signal as one input to review, alongside evidence provenance and consequences, rather than as the model certifying its own answer.[1]
What remains unproven—and what would change the assessment
The study does not test abstention, error correction after reflection, source citation, adversarial questions or uncertainty about tool outputs. Confidence targets partly originate from the same model's sampled behaviour, and semantic clustering for open answers uses another language model. Although factual calibration anchors the scores to correctness, the method can inherit benchmark artefacts and the sampled model's repeated misconceptions. Open code, responses and training records support inspection, but independent reproduction has not yet established generality.
The authors report US National Science Foundation support under grant IIS-2505006 and no competing interests. Experiments used OpenAI and Fireworks services, while the paper and public repository provide the datasets, generated responses and analysis code. The disclosed materials make the statistical analysis auditable, but they do not provide the proprietary base-model weights or reconstruct every service-side detail. Later API changes could produce different results even with the same prompts.
Confidence in the finding would rise if independent teams reproduced it across more model families, multilingual and multimodal tasks, retrieval and agent workflows, and prospective human-review studies. The safety assessment would change if multitask confidence training consistently improved real error detection without reducing answer quality or encouraging overreliance. For now, the strongest conclusion is narrower: models can learn better ways to express uncertainty, but training must match the form of judgment being used, and benchmark calibration is not evidence of safe professional deployment.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring large language models: compare the finding and its evidence limits.
Can compression make large language models practical on smaller hardware?
A peer-reviewed survey of 165 studies maps how pruning, quantisation, distillation and low-rank adaptation can reduce resource demands. It finds a toolkit, not one proven recipe—and benchmark differences obstruct fair comparison.
Read the next reportExplore the full source and evidence record
What this means for people
- Clinicians, lawyers and public-service staff should not treat a model's percentage as a guarantee about an individual answer; it is a population-level signal that needs local validation.
- Product teams can test whether a confidence display is calibrated after retrieval, tools and policy prompts are added, rather than relying on a base-model benchmark.
- Auditors and buyers can require results for the exact confidence operation a system exposes, because numeric reporting and pairwise ranking did not reliably transfer to one another.
Global context
The authors are based at the University of California, Irvine, and the training domains and prompts are in English. Medical and legal reasoning, professional duties and acceptable escalation thresholds differ across countries and languages. A confidence display validated on these benchmarks should not be exported into another jurisdiction or population without local cases, subgroup analysis and human-factors testing.
What the evidence does not yet show
- Only GPT-4.1-mini and Llama-3.1-70B-Instruct were tested, using model snapshots and API runs from 2025.
- The evaluation used benchmark questions, not deployed medical, legal, educational or public-service decisions.
- The consistency target came from ten samples of the same model and can be high for a repeatedly wrong answer; factual calibration mitigates but does not eliminate this problem.
- Generalization across domains was smaller and variable, while transfer between the two metacognitive tasks was unreliable.
- The study did not measure human interpretation, overreliance, abstention, tool-use uncertainty, multilingual performance or downstream harm.
What to watch next
- Independent replication across current proprietary and open-weight model families.
- Calibration after retrieval, web search, code execution and multi-step agent actions.
- Prospective studies measuring whether confidence signals improve human review rather than causing automation bias.
- Multilingual, subgroup and distribution-shift tests with preregistered thresholds.
- Comparisons with abstention, source-grounding and external verifier methods rather than confidence text alone.
Living evidence record
Impact record IAI-0WI56UT
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Whether supervised fine-tuning improves verbal uncertainty communication within and beyond the domains and metacognitive tasks used for training
Topics: Uncertainty · Calibration · Large language models · Fine-tuning · AI reliability
Regions: United States · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.