RLSR aligns LLMs with selective prediction using AURC

Researchers propose Reinforcement Learning for Selection Reward (RLSR), a novel framework that aligns large language models during post-training to optimize selective prediction performance. Unlike existing methods focusing on correctness or calibration, RLSR targets the area under the risk-coverage curve (AURC) as its alignment objective.

• The method allows LLMs to predict only for inputs where they are likely correct, reducing error rates and flagging uncertain cases for human review.
• RLSR achieves a substantially better risk-coverage trade-off compared to multiple alignment baselines on both in-domain and out-of-domain tasks.

This approach improves LLM reliability by balancing the risk-coverage trade-off and enabling seamless human-AI collaboration in high-stakes decision-making systems.