First-hand information for everyone
When should a language model answer a question, and when should it say “I don’t know”? A new study investigates that practical safety question by comparing two ways of deciding when a large language model should abstain from answering.
The team worked with Qwen3-14B, a 14-billion-parameter model, and focused on factual question-answering. They tested two broad approaches: (1) use the model’s own calibrated confidence scores and apply a threshold (a “confidence filter”), or (2) train a separate component — an answer selector — that learns to score the model’s candidate answers and decide whether to allow them.
Performance was measured with a single utility score that rewards correct answers, penalizes mistakes, and allows the system to abstain. This makes the trade-offs explicit: answering more often can increase coverage but also increase mistakes; abstaining avoids mistakes but reduces coverage.
On the TriviaQA dataset the learned selector produced clear gains. Averaged across three training runs and two held-out confidence thresholds, the selector raised mean coverage from 20.2% to 30.9% and cut the pooled error rate among answers that were allowed from 23.0% to 14.5% when answers were judged by exact-match criteria. A partial human-reviewed audit confirmed the improvement on TriviaQA, supporting the idea that training a separate selection model can be helpful in this setting.
The study also tried changing the base model itself using reinforcement learning to encourage better abstention behavior, but those interventions did not yield consistent improvements. That suggests it can be hard to get reliable abstention simply by fine-tuning the answerer with RL in these experiments.
On a second dataset, NQ-Open, results were less encouraging. The selector did better than the calibrated confidence filter on the averaged utility metric, but neither approach outperformed a simple “always abstain” baseline for the utility measure used. In other words, on that dataset the model’s answers were often not worth the risk of answering at all.
The authors draw cautious lessons. The success on TriviaQA supports the idea that a learned selection component can improve when a model should speak up. But the mixed results across datasets, the sensitivity to evaluation thresholds, and the remaining gap on NQ-Open underscore that reliable abstention across varied tasks and distributions is still an open problem. The findings also stress that evaluation choices — exact-match grading, how abstention is penalized, and dataset characteristics — matter a lot.
Implications for deployment: a learned selector can be a useful tool to reduce harmful or incorrect answers, but it should not be viewed as a full solution. Practitioners should test selection methods across many datasets, include human review when possible, and treat abstention policies cautiously under domain shift. Future work will need broader audits, improved out-of-distribution detection, and clearer ways to model the cost of abstaining versus answering.
In short, training a separate answer-selection component looks promising for making language models more selective about when to answer, but building generally reliable abstention remains an unresolved challenge.