Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection
A useful language model should know when to answer and when to say "I don't know.
ProofPaper ↗
Key points
- 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.
- Our results support learning to select answers in this setting while showing that reliable abstention across datasets remains unresolved.
Sources (1)
- [1]Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer SelectionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 04:33 PM
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?
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
- Oct 9, 2026SimVTA: Simulation-Grounded Video-Trajectory Alignment for Driving World Action Models
- Oct 9, 2026MENTOR: Learning from Failure with On-Policy Pedagogical Intervention
- Oct 9, 2026Rethinking Relative Policy Differences Transfer of On-Policy Distillation
- Oct 8, 2026Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents
- Oct 8, 2026SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference