Teaching a model to say "I don't know": a separate answer selector beats retraining the policy, lifting TriviaQA coverage from 20.2% to 30.9%

For short factual question answering, the decision of whether to release an answer can be handed to a selector separate from the generator, instead of changing the model that produces answers. The prior route let the original model supply its own confidence and filtered answers with a calibrated threshold; on TriviaQA that filter released only 20.2% of questions on average, and 23.0% of the answers it released were wrong. On 2,000 TriviaQA test questions at the two thresholds held out from policy training, 0.65 and 0.85, the selector raises mean coverage to 30.9% and lowers pooled error to 14.5%, with mean utility rising from 0.02418 to 0.04248, a paired improvement of 0.01830 and a 99% bootstrap interval of [0.01041, 0.02593], positive for each of the three training seeds.

The selector works by learning a separate scorer for whether a candidate answer is correct, then using it to decide between releasing the answer and returning IDK. The original model's own P(True) score only normalizes the next-token probability over the words True and False and is then calibrated by logistic regression; the selector instead uses a separate LoRA adapter that learns exact-match correctness from the question and candidate alone. The selector trains on 8,000 examples, one greedy plus three sampled candidates per training question from the original generator, and fits its own calibration map on 500 calibration questions; on 1,784 valid TriviaQA candidates, ROC AUC rises from 0.779 for the original score to 0.852 for the selector, and the calibrated Brier score falls from 0.189 to 0.154.

Giving the same training labels to a logistic-regression control does not recover the selector's gain, indicating the gain comes from learning new question-candidate features rather than from extra labels alone. That control applies two affine transformations to the original score, which compose into one affine transformation and cannot learn new question-candidate features. After calibration the selector's utility difference over that control is 0.01834 with a 99% interval of [0.01041, 0.02602], and both selector comparisons remain positive at the 99% level.

The results cover Qwen3-14B, short English factual question answering without retrieval, and three training seeds, with evaluation at thresholds 0.65 and 0.85 that were absent from policy training. Both the confidence filters and the selector add a scoring step after answer generation, and training and calibration are done on TriviaQA and transferred unchanged to NQ-Open without refitting. The intended readers are researchers and engineers working on selective prediction and abstention mechanisms.

The tested reinforcement-learning changes do not establish consistent gains: all three adjusted intervals include zero, and the direction of the differences varies across seeds. On NQ-Open the selector improves on the original calibrated filter but its mean utility stays below the zero of always abstaining, and the result at threshold 0.85 rests on a single released answer that is incorrect. The human review covers only 150 response units with AI-proposed labels and one human reviewer, not agreement between two independent annotators, so it cannot estimate the grading error rate over the full test sets.