Learning to Reason via Mixture-of-Thought for Logical Reasoning

2025-05-21

AI Summary

Existing large language model (LLM) approaches to logical reasoning predominantly rely on a single natural language modality, limiting their ability to synergistically integrate diverse reasoning paradigms and thereby constraining performance on complex logical problems. To address this, we propose the Mixture-of-Thoughts (MoT) framework, the first to introduce truth tables as a learnable symbolic reasoning modality—complementing natural language and code to form a tri-modal collaborative reasoning architecture. We further design a self-evolving multimodal distillation training mechanism that enables cross-modal chained reasoning and symbolic modeling. On benchmarks including FOLIO and ProofWriter, MoT achieves an average accuracy gain of 11.7 percentage points over state-of-the-art single-modality chain-of-thought methods, with particularly strong gains on higher-order logical reasoning tasks. Our core contributions are threefold: (1) the introduction of truth tables as a novel symbolic modality; (2) a tri-modal joint training paradigm; and (3) a self-evolving multimodal cooperative reasoning mechanism.

Technology Category

• Machine Learning: Multimodal Learning
• Knowledge Representation and Reasoning: Logic Programming
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Human beings naturally utilize multiple reasoning modalities to learn and solve logical problems, i.e., different representational formats such as natural language, code, and symbolic logic. In contrast, most existing LLM-based approaches operate with a single reasoning modality during training, typically natural language. Although some methods explored modality selection or augmentation at inference time, the training process remains modality-blind, limiting synergy among modalities. To fill in this gap, we propose Mixture-of-Thought (MoT), a framework that enables LLMs to reason across three complementary modalities: natural language, code, and a newly introduced symbolic modality, truth-table, which systematically enumerates logical cases and partially mitigates key failure modes in natural language reasoning. MoT adopts a two-phase design: (1) self-evolving MoT training, which jointly learns from filtered, self-generated rationales across modalities; and (2) MoT inference, which fully leverages the synergy of three modalities to produce better predictions. Experiments on logical reasoning benchmarks including FOLIO and ProofWriter demonstrate that our MoT framework consistently and significantly outperforms strong LLM baselines with single-modality chain-of-thought approaches, achieving up to +11.7pp average accuracy gain. Further analyses show that our MoT framework benefits both training and inference stages; that it is particularly effective on harder logical reasoning problems; and that different modalities contribute complementary strengths, with truth-table reasoning helping to overcome key bottlenecks in natural language inference.

Problem

• Enables LLMs to reason across multiple modalities (natural language, code, symbolic).
• Addresses limitations of single-modality training in logical reasoning tasks.
• Improves accuracy by leveraging complementary strengths of different reasoning modalities.

Innovation

• MoT integrates natural language, code, truth-table reasoning
• Self-evolving training with filtered multi-modal rationales
• Synergistic multi-modal inference enhances logical reasoning