Research map / Large language models / Reasoning and chain of thought

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

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Mixture-of-Thought (MoT), which self-evolves LLMs on filtered self-generated natural-language, code, and truth-table rationales and ensembles the three modalities by majority vote at inference, improving logical reasoning on FOLIO and ProofWriter.

Problem

LLM logical reasoning is typically trained and run in a single modality (natural language CoT), missing complementary strengths of code and symbolic formats; existing combinations act only at inference time.

Approach

Trains one model jointly on three reasoning modalities (NL CoT, code, and a new truth-table symbolic modality) via iterative on-policy self-generated and filtered rationales, then aggregates the three modalities by majority vote at inference.

Key results

• MoT training with single-thought inference raises average accuracy by up to 11.7pp over base models (e.g., Gemma-2-2B-It 41.1 to 61.9 avg).
• Gemma-2-9B-It MoT with mixture-of-thought inference reaches 78.9% on FOLIO, matching Logic-LM with GPT-4 (78.9%).
• Qwen2.5-7B-Instruct MoT achieves 75.1 average (78.3 FOLIO, 71.8 ProofWriter) with mixture inference.

Surprising findings

• Code-based reasoning has the lowest accuracy and oracle upper bound among the three modalities, even after MoT training.
• Random single-modality-per-question mixing performs far worse than direct mixing of all modalities (53.7 vs 62.6 on FOLIO).

Limitations

• Evaluated only on logical reasoning; not tested on math or commonsense reasoning.
• Truth-table modality is generated by the LLM itself rather than by an exact symbolic procedure, leaving room for improvement.
• Baselines like GPT-4 Logic-LM are cited from original papers rather than re-run under the same setup.

Experimental setup

Datasets: FOLIO ProofWriter ProverQA

Benchmarks: FOLIO ProofWriter ProverQA

Metrics: accuracy pass@k ensemble accuracy

Baselines: Logic-LM CoT Self-Consistency (@3) Single-Thought training Symbolic CoT LINC HybridMind Logic-of-Thoughts

Builds on: STaR Logic-LM Self-Consistency HybridMind vLLM

Closest papers by meaning

• RATT: A Thought Structure for Coherent and Correct LLM Reasoning (AAAI 2025 · similarity 0.90)
• SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond (NeurIPS 2025 · similarity 0.89)
• Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.89)
• MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning (ICLR 2024 · similarity 0.88)
• Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space (NeurIPS 2025 · similarity 0.88)
• Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.88)