Process Supervision of Confidence Margin for Calibrated LLM Reasoning

arXiv:2604.23333v1 Announce Type: cross Abstract: 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 (extbf{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.

Related

• Optimal Aggregation of LLM and PRM Signals for Efficient Test-Time Scaling
• ATTNPO: Attention-Guided Process Supervision for Efficient Reasoning
• Process Reward Models Meet Planning: Generating Precise and Scalable Datasets for Step-Level Rewards
• SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions
• Learning to Correct: Calibrated Reinforcement Learning for Multi-Attempt Chain-of-Thought
• Efficient PRM Training Data Synthesis via Formal Verification

Source: arXiv cs.CL | 2026-04-28