SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought

TL;DR

This paper introduces Flow Chain of Thought (Flow CoT), a new reasoning paradigm modeled as a progressive trajectory of latent states, and presents SCOUT, a lightweight fine-tuning framework that enhances reasoning in large language models without pretraining.

Contribution

The paper proposes Flow CoT as a scalable reasoning framework and introduces SCOUT, a novel fine-tuning method that improves reasoning accuracy and explanation quality without requiring pretraining.

Findings

SCOUT achieves up to 1.8% accuracy gains on reasoning benchmarks.

Flow CoT enables deeper, iterative reasoning and better explanation granularity.

SCOUT is effective across eight reasoning benchmarks.

Abstract

Chain of Thought (CoT) prompting improves the reasoning performance of large language models (LLMs) by encouraging step by step thinking. However, CoT-based methods depend on intermediate reasoning steps, which limits scalability and generalization. Recent work explores recursive reasoning, where LLMs reuse internal layers across iterations to refine latent representations without explicit CoT supervision. While promising, these approaches often require costly pretraining and lack a principled framework for how reasoning should evolve across iterations. We address this gap by introducing Flow Chain of Thought (Flow CoT), a reasoning paradigm that models recursive inference as a progressive trajectory of latent cognitive states. Flow CoT frames each iteration as a distinct cognitive stage deepening reasoning across iterations without relying on manual supervision. To realize this, we propose SCOUT (Stepwise Cognitive Optimization Using Teachers), a lightweight fine tuning framework that enables Flow CoT style reasoning without the need for pretraining. SCOUT uses progressive distillation to align each iteration with a teacher of appropriate capacity, and a cross attention based retrospective module that integrates outputs from previous iterations while preserving the models original computation flow. Experiments across eight reasoning benchmarks show that SCOUT consistently improves both accuracy and explanation quality, achieving up to 1.8% gains under fine tuning. Qualitative analyses further reveal that SCOUT enables progressively deeper reasoning across iterations refining both belief formation and explanation granularity. These results not only validate the effectiveness of SCOUT, but also demonstrate the practical viability of Flow CoT as a scalable framework for enhancing reasoning in LLMs.