From Chaos to Order: The Atomic Reasoner Framework for Fine-grained Reasoning in Large Language Models

TL;DR

Atomic Reasoner (AR) introduces a slow-thinking framework that decomposes reasoning into Atomic Reasoning Actions (ARAs) and uses a Cognitive Routing over an Atomic Tree to guide LLMs without additional training. By separating higher-level guidance from execution and applying domain-adaptable SOPs and a rigorous checker, AR aims to reduce cognitive load while maintaining coherence in long-sequence reasoning. Empirical results show strong gains on linguistic logic tasks and robust performance across baselines, with AR-generated data enhancing downstream fine-tuning. The work presents a modular, scalable approach to improving deliberative reasoning in LLMs, highlighting potential for cross-domain application and future integration with reinforcement learning.

Abstract

Recent advances in large language models (LLMs) have shown remarkable progress, yet their capacity for logical ``slow-thinking'' reasoning persists as a critical research frontier. Current inference scaling paradigms suffer from two fundamental constraints: fragmented thought flows compromising logical coherence, and intensively computational complexity that escalates with search space dimensions. To overcome these limitations, we present \textbf{Atomic Reasoner} (\textbf{AR}), a cognitive inference strategy that enables fine-grained reasoning through systematic atomic-level operations. AR decomposes the reasoning process into atomic cognitive units, employing a cognitive routing mechanism to dynamically construct reasoning representations and orchestrate inference pathways. This systematic methodology implements stepwise, structured cognition, which ensures logical coherence while significantly reducing cognitive load, effectively simulating the cognitive patterns observed in human deep thinking processes. Extensive experimental results demonstrate AR's superior reasoning capabilities without the computational burden of exhaustive solution searches, particularly excelling in linguistic logic puzzles. These findings substantiate AR's effectiveness in enhancing LLMs' capacity for robust, long-sequence logical reasoning and deliberation.

Figures (6)

• Figure 1: Illustration of the analysis of the slow thinking process for o1, with the case from the OpenAI website.
• Figure 2: Overall architecture of Atomic Reasoner (AR). AR introduces Atomic Reasoning Actions and proposes an Atomic Tree-based organization of reasoning processes. Unlike traditional Tree-Based Expansion methods (top right) that rely on LLM's parallel node expansion and complex selections, AR adopts a cognitive reasoning approach, guiding LLMs to first select appropriate reasoning actions, followed by precise operations (node expansion, backtracking, or termination) on the Atomic Tree, ultimately achieving more efficient deliberate reasoning.
• Figure 3: Schematic illustration of error types defined in the Checker mechanism of AR.
• Figure 4: Ablation studies on different mechanisms in AR, with results from model GLM-4-flashX (left) and model GPT-4o-mini (right).
• Figure 5: Performance curves with respect to the scaling of Maximum Reasoning Rounds.

Theorems & Definitions (1)

• Remark 1