Biomedical fine-tuning often reduced AI models’ caution in paired audit

Testing when not to answer

The study examines what the authors call “knowledge-centric abstention”: withholding an answer because the available evidence is insufficient. This differs from refusing an unethical or harmful request. In medicine, the relevant question is whether a system distinguishes between having enough information to answer and needing more.

The researchers compared 25 publicly released biomedical models with the 11 parent checkpoints from which they descended. The models spanned Llama 2, Llama 3, Llama 3.1 and Gemma 3, with varied training methods. This was an observational audit of existing releases, not an experiment that held training conditions fixed.

Two benchmarks tested different aspects of caution. MedAbstain supplied 1,007 medical multiple-choice questions in each of four versions: original or evidence-removed questions, each with or without an explicit abstention option. The evidence-removed version offering abstention was the main test of appropriate caution.

ClinDet-Bench tested whether clinical scoring criteria were met, not met or impossible to determine. Its 94 scenarios included 32 with complete information, 30 with missing information that did not prevent a determination, and 32 where missing information did prevent one.

All models received common plain-text prompts and generated responses using greedy decoding, which selects the highest-probability next token. The analysis compared changes within matched model pairs. Statistical checks accounted for multiple comparisons and for several specialized models sharing the same parent.

Less abstention, but not in every model

On MedAbstain’s evidence-removed questions, appropriate abstention decreased in 21 of 25 pairs and increased in four. Median abstention levels were 8.3% for parent models and 4.2% for specialized models. The median change calculated within pairs was a decline of 3.48 percentage points; this differs from subtracting the two group medians.

That decline remained statistically significant after the authors’ multiple-testing correction and shared-parent analysis. ClinDet-Bench pointed in the same direction, with a median decline of 6.25 percentage points on indeterminable cases, but its result did not meet those robustness criteria. The strongest evidence for reduced caution therefore comes from MedAbstain, not both benchmarks equally.

Less abstention could be useful if models became better at recognizing which questions remained answerable. The researchers tested this by comparing abstention on answerable and insufficient-evidence conditions. They found no reliable improvement in that distinction. They also did not establish a robust deterioration in this ability across the cohort.

Accuracy and caution did not follow one universal relationship. Sixteen pairs improved accuracy on MedAbstain’s original, forced-answer questions while becoming less cautious on evidence-removed questions. One improved both measures. No examined training-method group or model-size measure reliably explained the variation after statistical correction.

Newly answered questions were harder

A separate analysis examined original, answerable questions where abstention was optional. Questions that the specialized model answered but its parent declined were comparatively difficult: the specialized model’s accuracy on them was a median 14.98 percentage points below its accuracy on questions both models answered.

This does not show that adaptation erased knowledge. The groups were defined partly by the parent’s decision to abstain, so newly answered questions were already likely to be harder. When both models were required to answer the same underlying questions, no accuracy change within these groups survived multiple-testing correction.

The models also concentrated more probability on their preferred answers. On original MedAbstain questions without abstention, maximum-choice probability increased by a median 0.127. On questions both models answered incorrectly, it increased by 0.112. Greater commitment therefore extended to shared errors, rather than simply reflecting more correct answers. Confidence did not reliably improve at distinguishing correct from incorrect predictions.

These measurements concern probabilities normalized across the available answer choices—not confidence in an entire generated explanation. A smaller full-vocabulary check preserved the direction of shared-error changes, although normalization affected their magnitude.

What the audit cannot establish

MedAbstain’s evidence removal is imperfect. In its reported human-validation sample, abstention was considered appropriate in only 38 of 49 perturbed cases. Item-level validation labels were unavailable for this audit. Many models also had confirmed or possible exposure to MedQA-derived training material, creating a contamination risk that sensitivity analyses could not eliminate.

ClinDet-Bench offered rule-based labels but relatively few cases. Neither benchmark tested conversations, requests for additional information or escalation to a clinician. A nine-pair check using native chat templates reproduced the direction of reduced abstention but was not independently statistically conclusive; publisher-recommended system prompts remained untested.

The authors recommend evaluating accuracy, appropriate abstention and confidence together, against each model’s parent. Their evidence supports that broader audit—not a claim that all biomedical fine-tuning damages safety, nor a judgment that any tested model is ready for clinical deployment.