Mixture-of-Thought (MoT) for Logical Reasoning

Learning to Reason via Mixture-of-Thought for Logical Reasoning
Tong Zheng*, Lichang Chen*, Simeng Han, R. Thomas McCoy, Heng Huang
* These authors contributed equally

This repository contains code and scripts for the Mixture-of-Thought (MoT) framework, which combines natural-language CoT, code-based CoT, and symbolic truth-table CoT in a self-evolving training loop and a voting-based inference strategy to improve performance on logical reasoning benchmarks (FOLIO, ProofWriter).

---

🔥 Key Features

• Self-Evolving MoT Training
  • Multi-round fine-tuning: generate reasoning traces in three modalities, filter and verify them, then fine-tune for the next round.
• Mixture-of-Thought Inference
  • Simultaneously run NL-CoT, Code-CoT, and Truth-Table-CoT, then fuse answers by majority vote.
• Truth-Table-CoT
  • Identifies Variables, constructs partial truth table where each row does not conflict with premises, and checks conclusion consistency across surviving rows.
• Reproducible & Extensible
  • Tested with Gemma-2-2B-It, Gemma-2-9B-It, Qwen-2.5-7B-Instruct, and easily adaptable to other LLMs.

---

🔥 Our MoT Framework

Figure: Overview of the Mixture-of-Thought (MoT) framework combining NL-CoT, Code-CoT, and Truth-Table-CoT in a self-evolving training loop and voting-based inference.

---

🔥 Main Results

Model	Method Type	Reasoning Modality	FOLIO (%)	ProofWriter (%)	Avg (%)
Prior SOTA: GPT-4 (Logic-LM)	Logic-LM	-	78.9	79.7	79.3
Prior SOTA: GPT-4 (CoT)	CoT (Vanilla)	-	70.6	68.1	69.4
Gemma-2-2B-It (3-Shot)	Single-Thought	Best (nl)	42.4	39.8	41.1
Gemma-2-2B-It @3 (3-Shot)	Single-Thought	Best (nl)	45.3	38.8	42.1
MoT (0-Shot)	Single-Thought	Best	61.1	62.7	61.9
MoT (0-Shot)	Mixture-of-Thought	All	62.6	65.0	63.8
Gemma-2-9B-It (3-Shot)	Single-Thought	Best (nl)	69.5	61.2	65.4
Gemma-2-9B-It @3 (3-Shot)	Single-Thought	Best (nl)	72.9	62.7	67.8
MoT (0-Shot)	Single-Thought	Best	76.9	69.5	73.2
MoT (0-Shot)	Mixture-of-Thought	All	78.9	70.7	74.8
Qwen2.5-7B-Instruct (3-Shot)	Single-Thought	Best (nl)	71.9	60.5	66.2
Qwen2.5-7B-Instruct @3 (3-Shot)	Single-Thought	Best (nl)	73.4	65.8	69.6
MoT (0-Shot)	Single-Thought	Best	75.9	69.2	72.6
MoT (0-Shot)	Mixture-of-Thought	All	78.3	71.8	75.1

🚀 Quick Start

1. Environment Setup

# Create and activate a conda environment
conda env create -f environment.yml

2. Self-evolving Mixture-of-Thought Training

# Self-evolving Training with on Policy 
bash scripts/star_pipeline_final_on_policy.sh (On Policy)

# Self-evolving Mixture-of-Thought Training with on Policy 
bash scripts/mix_training_gemma_2_9b.sh (On Policy)

# Self-evolving Mixture-of-Thought Training with off Policy 
bash scripts/mix_training_gemma_2_9b_offp.sh (Off Policy)

3. Mixture-of-Thought Inference

# Inference with each thought paradigm 
bash scripts/eval.sh # set MODE="nl"
bash scripts/eval.sh # set MODE="truth_table"
bash scripts/eval.sh # set MODE="code"

# Majority Vote
python analysis/hard_vote.py [file_1] [file_2] [file_3] 

4. Experiments: Mixture-of-Thought Sampling vs. Single-thought Sampling

# Sample 128 outputs for each thought paradigms and each model (baseline and our MoT model)
bash scripts/budget_experiments/eval.sh

# Perform Mixture-of-Thought sampling and Single-thought Sampling from these outputs
python budget_experiment/eval_test_time_scaling_vote.py (if you want to plot accuracy curve)

python budget_experiment/eval_test_time_scaling.py (if you want to plot pass@k curve)

5. Experiments: Performance across Problem Difficulties

# Single Thought for FOLIO (Hint: you do not need to run this on ProverQA as they have provide separate subset with different difficulties)
python difficulty_experiments/eval_by_difficulty.py --prediction_path [prediction_path] --easy_dataset_name [easy_dataset_name] --difficult_dataset_name [difficult_dataset_name]

# Mixture-of-Thought for FOLIO
python difficulty_experiments/eval_vote_by_difficulty.py --nl_path [nl_path] --code_path [code_path] --tt_path [tt_path] --easy_dataset_name [easy_dataset_name] --difficult_dataset_name [difficult_dataset_name]

6. Complementarity & Uniqueness Analysis

cd analysis

python analysis_overlap.py [file_1] [file_2] [file_3]

# The files file_1, file_2, and file_3 contain the inference outputs for each reasoning paradigm and can be generated using our evaluation scripts.

📖 Citation

If you use this code or the MoT methodology, please cite:

@article{zheng2025learning,
  title={Learning to Reason via Mixture-of-Thought for Logical Reasoning},
  author={Zheng, Tong and Chen, Lichang and Han, Simeng and McCoy, R Thomas and Huang, Heng},
  journal={arXiv preprint arXiv:2505.15817},
  year={2025}
}