Parallel thinking enhances the reasoning ability of large language models (LLMs) by exploring multiple reasoning paths simultaneously. However, reinforcement‑learning‑based parallel thinking often suffers from mode collapse, where different trajectories become similar, eroding hypothesis diversity and leading to premature convergence on suboptimal solutions. To address this, we introduce Diverse Parallel Thinking (DPT), a simple yet effective RL approach that penalizes prefix similarity among parallel paths, encouraging distinct reasoning strategies and a more exhaustive exploration of the reasoning space. Experiments on several reasoning benchmarks show that DPT outperforms previous parallel‑thinking methods and achieves state‑of‑the‑art performance over baselines. The results confirm that path diversity is a crucial factor for improving parallel thinking. The code is released at https://github.com/JiaruQian/DPT.
Review