Mixture of Reasonings: Teach Large Language Models to Reason with Adaptive Strategies

2025-07-01

AI Summary

To address the limitations of large language models (LLMs)—including reliance on manually engineered task prompts, poor generalization across reasoning tasks, and suboptimal inference efficiency—this paper proposes the Mixture of Reasoning (MoR) framework. MoR is the first approach to internalize diverse reasoning strategies directly into model parameters, enabling prompt-free, task-adaptive reasoning without external prompting engineering. Methodologically, it employs a two-stage data construction pipeline: (1) high-quality chain-of-thought (CoT) templates are generated using GPT-4o, and (2) these templates are paired with benchmark tasks for supervised fine-tuning (SFT). Evaluated under standard CoT prompting, MoR150 achieves 0.730 accuracy on multi-task reasoning benchmarks—outperforming strong baselines by 13.5 percentage points in relative gain (absolute improvement of +2.2%). This demonstrates substantial gains in cross-task generalization and reasoning robustness.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Search and Optimization: Metareasoning and Metaheuristics
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large language models (LLMs) excel in complex tasks through advanced prompting techniques like Chain-of-Thought (CoT) and Tree-of-Thought (ToT), but their reliance on manually crafted, task-specific prompts limits adaptability and efficiency. We introduce Mixture of Reasoning (MoR), a training framework that embeds diverse reasoning strategies into LLMs for autonomous, task-adaptive reasoning without external prompt engineering. MoR has two phases: Thought Generation, creating reasoning chain templates with models like GPT-4o, and SFT Dataset Construction, pairing templates with benchmark datasets for supervised fine-tuning.Our experiments show that MoR significantly enhances performance, with MoR150 achieving 0.730 (2.2% improvement) using CoT prompting and 0.734 (13.5% improvement) compared to baselines. MoR eliminates the need for task-specific prompts, offering a generalizable solution for robust reasoning across diverse tasks.

Problem

• Enables LLMs to autonomously adapt reasoning strategies without manual prompts
• Integrates diverse reasoning methods into a single training framework
• Improves performance across tasks by eliminating task-specific prompt engineering

Innovation

• Embed diverse reasoning strategies into LLMs
• Autonomous task-adaptive reasoning without prompts
• Enhance performance with supervised fine-tuning