S3-CoT: Self-Sampled Succinct Reasoning Enables Efficient Chain-of-Thought LLMs

TL;DR

CoT traces can improve reasoning but often become verbose and costly. S$^{3}$-CoT introduces a self-sampling pipeline that uses activation steering to identify a variable-length direction (VL-D) and sample CoTs directly from target LLMs, followed by verification and training with a dual-cognitive, progressively compressing SFT strategy. The approach achieves superior or competitive length-accuracy trade-offs across math and cross-domain medical benchmarks, with strong adaptability across general and R1-style LLMs and reduced reliance on external supervision. It also explores an autonomous, self-evolving variant (S$^{3}$-CoT$^{sc}$), underscoring potential for LLMs to internalize efficient reasoning patterns. Limitations remain for some R1-style models, and future work aims to push beyond the current Pareto frontier between CoT length and accuracy.

Abstract

Large language models (LLMs) equipped with chain-of-thought (CoT) achieve strong performance and offer a window into LLM behavior. However, recent evidence suggests that improvements in CoT capabilities often come with redundant reasoning processes, motivating a key question: Can LLMs acquire a fast-thinking mode analogous to human System 1 reasoning? To explore this, our study presents a self-sampling framework based on activation steering for efficient CoT learning. Our method can induce style-aligned and variable-length reasoning traces from target LLMs themselves without any teacher guidance, thereby alleviating a central bottleneck of SFT-based methods-the scarcity of high-quality supervision data. Using filtered data by gold answers, we perform SFT for efficient CoT learning with (i) a human-like dual-cognitive system, and (ii) a progressive compression curriculum. Furthermore, we explore a self-evolution regime in which SFT is driven solely by prediction-consistent data of variable-length variants, eliminating the need for gold answers. Extensive experiments on math benchmarks, together with cross-domain generalization tests in medicine, show that our method yields stable improvements for both general and R1-style LLMs. Our data and model checkpoints can be found at https://github.com/DYR1/S3-CoT.

Figures (3)

• Figure 1: A self-sampling framework for efficient CoT learning. Our study (1) samples variable-length CoT data via intervention along VL-D; (2) filters data via answer or self-consistency verification; and (3) achieves efficient CoT internalization via a dual-cognitive system and progressive compression curriculum.
• Figure 2: Analysis of VL-D properties. We provide PCA-based visualizations and quantify how the mean separation strength and angle variance metric vary across layers. Visualizations across all layers under various LLMs are in Fig. \ref{['app_vis_qw2.5']}, \ref{['app_vis_dsqw']}, \ref{['app_vis_ll3']}, and \ref{['app_vis_qw3']}, respectively. Analysis on LLaMA3$_{8B}$ and Qwen3-Think$_{4B}$ are in Fig. \ref{['app_fig_ana']} .
• Figure 3: Probe experiments on intervention layers and strength. Green: average Len-R; Yellow: number of collapsed samples; Green "×": all samples collapse. Bottom-right: Len-R distribution under large-scale sampling. Results for LLaMA3$_{8B}$ and Qwen3-Think$_{4B}$ are in Fig. \ref{['app_fig_ratio']}, and results for other intervention settings are in Fig. \ref{['app_interve_other_sets']}.