Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Process Supervision of Confidence Margin for Calibrated LLM Reasoning

About

Scaling test-time computation with reinforcement learning (RL) has emerged as a reliable path to improve large language models (LLM) reasoning ability. Yet, outcome-based reward often incentivizes models to be overconfident, leading to hallucinations, unreliable confidence-based control, and unnecessary compute allocation. We introduce Reinforcement Learning with Confidence Margin (\textbf{RLCM}), a calibration-aware RL framework that jointly optimizes correctness and confidence reliability via a margin-enhanced process reward over intermediate-budget completions. Rather than aligning confidence to correctness likelihoods, RLCM encourages to widen the confidence margin between correct and incorrect steps within a single reasoning trajectory. Across mathematical, code, logic and science benchmarks, our method substantially improves calibration while maintaining or improving accuracy. We further show that, with calibrated confidence signals, the resulting models enable more efficient conformal risk control and effective confidence-weighted aggregation.

Liaoyaqi Wang, Chunsheng Zuo, William Jurayj, Benjamin Van Durme, Anqi Liu• 2026

Related benchmarks

TaskDatasetResultRank
ReasoningReasoning Suite Average
Accuracy74.8
45
Logical reasoningLogiQA
Accuracy49.7
34
CodingLiveCodeBench
Accuracy (Acc)38.9
5
Mathematical ReasoningAIME 24
Accuracy49.2
5
Science ReasoningGPQA
Accuracy37.1
5
Mathematical ReasoningAIME 25
Accuracy36
5
Mathematical ReasoningMATH 500
Accuracy89.1
5
Mathematical ReasoningAMC
Accuracy78.5
5
Mathematical ReasoningIn-domain math benchmarks Overall
Accuracy61.8
5
Mathematical ReasoningOlympiadBench
Accuracy56
5
Showing 10 of 10 rows

Other info

Follow for update