aiqbench analysis
This research investigates whether language models can be effectively trained to distinguish between answering and abstaining, finding that a learned selector improves utility on factual datasets like TriviaQA compared to standard confidence filters. However, the study highlights that reliable abstention remains an unresolved challenge across diverse benchmarks, with reinforcement learning approaches failing to yield consistent gains. The work underscores the difficulty of balancing coverage and accuracy in open-ended QA tasks without degrading performance on specific domains.
Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
arXiv:2610.13821v1 Announce Type: new Abstract: A useful language model should know when to answer and when to say "I don't know." Can we improve that decision by training the model itself or by training a separate component to check its answers? We study both approaches on Qwen3-14B for factual questions, using a score that rewards correct answers, penalizes mistakes, and allows abstention. Across three training seeds and two held-out confidence thresholds, a learned answer selector improves threshold-averaged utility over the original model's calibrated confidence filter. On TriviaQA, mean coverage increases from 20.2% to 30.9%, while the pooled error rate among scored answers decreases from 23.0% to 14.5% under exact-match grading. The tested reinforcement-learning changes do not establish consistent gains. On NQ-Open, the selector improves threshold-averaged utility over the filter but still performs worse than always abstaining. A partial human-reviewed audit preserves the TriviaQA finding. Our results support learning to select answers in this setting while showing that reliable abstention across datasets remains unresolved.
Title:Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
Abstract:A useful language model should know when to answer and when to say "I don't know." Can we improve that decision by training the model itself or by training a separate component to check its answers? We study both approaches on Qwen3-14B for factual questions, using a score that rewards correct answers, penalizes mistakes, and allows abstention. Across three training seeds and two held-out confidence thresholds, a learned answer selector improves threshold-averaged utility over the original model's calibrated confidence filter. On TriviaQA, mean coverage increases from 20.2% to 30.9%, while the pooled error rate among scored answers decreases from 23.0% to 14.5% under exact-match grading. The tested reinforcement-learning changes do not establish consistent gains. On NQ-Open, the selector improves threshold-averaged utility over the filter but still performs worse than always abstaining. A partial human-reviewed audit preserves the TriviaQA finding. Our results support learning to select answers in this setting while showing that reliable abstention across datasets remains unresolved.
Computation and Language (cs.CL)
arXiv:2610.13821 [cs.CL]
(or arXiv:2610.13821v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.13821 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 266 words
Related coverage
Hierarchical Group-Conditional Conformal Risk Control for Selective Prediction in Language Models
arXiv · Jul 28, 2026
Causal Evidence that Language Models use Confidence to Drive Behavior
arXiv · Sep 21, 2026
From Probe Scores to Alarm Policies: Operational Validity of Activation Monitors for Language-Model Agents
arXiv · Oct 7, 2026
Visual Grounding in Zero-Shot Vision-Language Control
arXiv · Aug 7, 2026