Divergent-Convergent Reasoning Boosts LLM Accuracy with Less Compute
A new arXiv paper (2608.15303) introduces Divergent-Convergent Reasoning (DCR), a two-phase method for improving Large Language Model (LLM) reasoning. The approach first generates multiple candidate solutions (divergent phase) and then reconciles them (convergent phase). The authors report three main findings: a single reconciliation step can amplify correct minority answers, outperforming majority voting in cases where correct outputs are rare; recursive DCR, which iteratively analyzes disagreements and allocates compute, achieves higher accuracy than fixed-compute baselines; and on AIME 2024 and AIME 2025 benchmarks, recursive DCR reaches 93.3% and 92.0% accuracy respectively, while using about 27% less compute. The paper is available on arXiv and was announced as a new submission.
Key facts
- Paper arXiv:2608.15303 introduces Divergent-Convergent Reasoning (DCR).
- DCR consists of an exploration phase generating multiple solutions and a reconciliation phase.
- A single reconciliation step can amplify correct minority reports.
- DCR often recovers correct answers when correct outputs are in the minority, where majority voting fails.
- Recursive DCR iteratively analyzes disagreements and allocates test-time compute.
- Recursive DCR achieves 93.3% on AIME 2024 and 92.0% on AIME 2025.
- Recursive DCR uses roughly 27% less compute than fixed-compute baselines.
- The paper is announced as a new submission on arXiv.
Entities
Institutions
- arXiv