Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
A learned answer selector beat the model’s own confidence filter on TriviaQA, but abstention did not generalize cleanly.
The paper tests Qwen3-14B on factual QA with a utility score that rewards correct answers, penalizes wrong ones, and permits “I don’t know.” The selector raised TriviaQA coverage from 20.2% to 30.9% while lowering pooled exact-match error from 23.0% to 14.5%. Reinforcement-learning changes did not show consistent gains. On NQ-Open, the selector improved over the confidence filter but still lost to always abstaining. ArXiv · AI/CL/LG's note
The paper tests Qwen3-14B on factual QA with a utility score that rewards correct answers, penalizes wrong ones, and permits “I don’t know.” The selector raised TriviaQA coverage from 20.2% to 30.9% while lowering pooled exact-match error from 23.0% to 14.5%. Reinforcement-learning changes did not show consistent gains. On NQ-Open, the selector improved over the confidence filter but still lost to always abstaining. ArXiv · AI/CL/LG's note
score 5