D-COT: Disciplined Chain-of-Thought Learning for Efficient Reasoning in Small Language Models

TL;DR

This paper introduces D-CoT, a structured reasoning framework for small language models that improves accuracy and efficiency by guiding reasoning processes with control tags during training.

Contribution

D-CoT is a novel method that enforces disciplined reasoning in small language models using control tags, reducing overthinking and computational costs.

Findings

Significant accuracy improvements on GPQA-diamond and MMLU-Pro benchmarks.

Reduces token consumption while enhancing reasoning performance.

Models internalize structured reasoning, performing well without control tags during inference.

Abstract

Chain-of-Thought (CoT) distillation from Large Language Models (LLMs) often induces "overthinking" in Small Language Models (SLMs), leading to performance degradation and excessive token consumption. In this study, we propose Disciplined Chain-of-Thought (D-CoT), a novel framework that enforces a structured reasoning process using control tags -- such as <TEMP_LOW> for fact-checking and <TEMP_HIGH> for multi-perspective exploration -- as auxiliary scaffolding during training. By optimizing the CoT trajectory, D-CoT suppresses reasoning drift and simultaneously achieves token reduction and performance improvement. We demonstrate the efficacy of our approach on Qwen3-8B: with only 5,000 training samples, D-CoT significantly boosts accuracy on GPQA-diamond by 9.9% and MMLU-Pro (0-shot) by 9.1%, while drastically reducing computational costs. Furthermore, we confirm that the model internalizes this disciplined thought structure, maintaining high performance even without explicit control tags during inference.