Blind Spot Navigation in Large Language Model Reasoning with Thought Space Explorer

Existing structured reasoning methods improve large language models by extending or comparing previously generated reasoning paths, but they may repeatedly explore similar directions while overlooking other promising regions of the solution space. Thought Space Explorer (TSE) addresses these blind spots by identifying high-impact nodes, integrating information from multiple reasoning chains to create new nodes, and extending new branches through connection strategies. Experiments on mathematical reasoning and question-answering benchmarks show that TSE improves both final-answer accuracy and intermediate reasoning quality while maintaining a favorable effectiveness-efficiency trade-off.

📄 In Findings of the Association for Computational Linguistics: EACL 2026, pp. 3691–3707.
🔗 Paper Link (ACL Anthology)