# PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chains

**Authors:** Joshua Ong Jun Leang, Zheng Zhao, Aryo Pradipta Gema, Sohee Yang, Wai-Chung Kwan, Xuanli He, Wenda Li, Pasquale Minervini, Eleonora Giunchiglia, Shay B. Cohen  
**arXiv:** [2508.21787](https://arxiv.org/abs/2508.21787)  
**Date:** 2025-08-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.21787v1/

## Key points

- PiCSAR is a method designed to improve the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward.
- The key challenge in reasoning tasks is developing a scoring function that can identify correct reasoning chains without access to ground-truth answers.
- PiCSAR proposes a training-free approach that scores each candidate generation based on the joint log-likelihood of the reasoning process and the final answer, breaking it down into reasoning confidence and answer confidence.
- PiCSAR has shown significant improvements across various benchmarks, outperforming baseline methods with fewer samples in most comparisons.
- Correct reasoning chains exhibit higher levels of both reasoning and answer confidence, validating the effectiveness of PiCSAR.
- Higher peak-to-sentence ratios correspond to higher accuracy rates across different models, indicating that correct reasoning chains tend to have higher information density.
- Longer reasoning chains do not necessarily lead to improved accuracy; longer responses are often less accurate compared to shorter ones.

## Summary

, , , , 

PiCSAR, or Probabilistic Confidence Selection and Ranking for Reasoning Chains, is a method designed to improve the accuracy of large language models (LLMs) and large reasoning models (LRMs). It achieves this by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge in reasoning tasks is developing a scoring function that can identify correct reasoning chains without access to ground-truth answers. In response to this challenge, PiCSAR proposes a simple, training-free approach that scores each candidate generation based on the joint log-likelihood of the reasoning process and the final answer. The joint log-likelihood of the reasoning and final answer can be broken down into reasoning confidence and answer confidence. Through its implementation, PiCSAR has shown significant improvements across various benchmarks, outperforming baseline methods with fewer samples in most comparisons. An analysis of correct reasoning chains reveals that they exhibit higher levels of both reasoning and answer confidence, validating the effectiveness of PiCSAR. Furthermore, a study on sentence-level confidence dynamics as a proxy for reasoning quality sheds light on how PiCSAR operates. By analyzing how model confidence in the final answer evolves throughout a reasoning chain composed of multiple sentences, insights are gained into the relationship between confidence levels and accuracy. The findings suggest that higher peak-to-sentence ratios correspond to higher accuracy rates across different models, indicating that correct reasoning chains tend to have higher information density. Additionally, it is observed that longer reasoning chains do not necessarily lead to improved accuracy. In fact, longer responses are often less accurate compared to shorter ones. This observation aligns with recent research on inverse scaling in test-time compute. In conclusion, PiCSAR offers a promising approach to enhancing the performance of large language and reasoning models by leveraging probabilistic confidence selection and ranking mechanisms. Its ability to identify correct reasoning chains based on confidence levels contributes significantly to improving overall accuracy in diverse benchmark tasks.