# SEAL: Steerable Reasoning Calibration of Large Language Models for Free

**Authors:** Runjin Chen, Zhenyu Zhang, Junyuan Hong, Souvik Kundu, Zhangyang Wang  
**arXiv:** [2504.07986](https://arxiv.org/abs/2504.07986)  
**Date:** 2025-04-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.07986v1/

## Key points

- Large Language Models (LLMs) like OpenAI's o1-series excel in complex reasoning tasks through extended chain-of-thought (CoT) mechanism
- Studies reveal significant redundancy in CoT reasoning traces, leading to increased inference latency and decreased model performance
- LLMs' internal reasoning structures categorized into execution thoughts, reflection thoughts, and transition thoughts
- Excess of reflection and transition thoughts linked to failure cases with clear separation in latent space
- SEAL (Steerable Reasoning Calibration) introduced as training-free method to calibrate CoT process using steering vector in latent space
- SEAL demonstrated high transferability across tasks with up to 11% accuracy improvement and reduced reasoning tokens by 11.8% to 50.4%
- Fine-grained analysis of LLMs' CoT processes revealed inefficiency due to excessive reflection and transition thoughts leading to computational overhead
- Focus on developing controllable approach to mitigate redundant reflection and transition thoughts for improved efficiency and effectiveness of LLM reasoning processes

## Summary

In recent years, Large Language Models (LLMs) such as OpenAI's o1-series have demonstrated impressive capabilities in handling complex reasoning tasks through the extended chain-of-thought (CoT) reasoning mechanism. However, studies have uncovered significant redundancy in the CoT reasoning traces, resulting in increased inference latency and decreased model performance due to unnecessary diversion of attention towards irrelevant reasoning paths. To address this issue, a detailed analysis was conducted on the internal reasoning structures of LLMs, categorizing them into three primary thought types: execution thoughts for problem-solving step-by-step analysis, reflection thoughts for verification pauses during reasoning, and transition thoughts for shifting perspectives in problem-solving flow. The study revealed that an excess of reflection and transition thoughts were strongly associated with failure cases, with clear separation observed in the latent space among these thought categories. Building upon these findings, a novel approach named SEAL (Steerable Reasoning Calibration) was introduced as a training-free method to effectively calibrate the CoT process. SEAL involves an offline stage for extracting the steering vector in the latent space and an on-the-fly calibration of the reasoning trace through representation intervention using this vector. Notably, the steering vector demonstrated high transferability across various tasks. Extensive experiments were conducted across multiple models (DeepSeek-R1-Distill and QwQ-32B-Preview) and benchmarks (Math500, GSM8K, LiveCodeBench), validating SEAL's effectiveness. The results showed up to an 11% improvement in accuracy while reducing reasoning tokens by 11.8% to 50.4%. The code for SEAL is publicly available on GitHub. Further investigation delved into analyzing fine-grained reasoning patterns of LLMs utilizing CoT processes by segmenting generated output into interconnected thoughts categorized as execution thoughts for step-by-step problem-solving analysis, reflecting thoughts for verification pauses during reasoning, and transition thoughts for shifting perspectives in problem-solving flow. Statistical analysis revealed that incorrect samples exhibited higher numbers of thoughts compared to correct ones due to excessive reflection and transition steps introducing redundancy beyond necessary reasoning processes. The study highlighted two major flaws in current LLM reasoning processes: efficiency concerns arising from frequent reflection and transition thoughts consuming significant token budgets leading to computational overhead; effectiveness issues stemming from distraction caused by these unnecessary thoughts resulting in suboptimal performance due to deviation from essential reasoning paths. Moving forward, efforts are focused on analyzing different thought roles within the latent space and developing a controllable approach to mitigate redundant reflection and transition thoughts for enhancing both efficiency and effectiveness of LLM reasoning processes.