Reproduce “ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Compute” on OpenResearch

About this paper

Hao Wen, Yifan Su, Feifei Zhang, Yunxin Liu, Yunhao Liu, Ya-Qin Zhang, Yuanchun Li

ParaThinker introduces a native parallel thinking paradigm for Large Language Models, enabling them to explore multiple reasoning paths concurrently and synthesize them for improved accuracy. It achieves substantial performance gains on mathematical reasoning benchmarks with negligible inference latency, particularly enhancing smaller models.

Key insights

• The 'Tunnel Vision' phenomenon, where early imperfect tokens irrevocably commit an LLM to a suboptimal reasoning path, is a primary bottleneck for sequential test-time compute scaling.
• Native parallel thinking (width scaling) is a more effective and efficient strategy than purely sequential (depth) scaling for enhancing LLM reasoning capabilities.
• The integration of parallel thought processes and summarization directly within the LLM's forward pass, facilitated by specialized tokens and embeddings, is critical for efficient and superior performance without external verifiers.

Summary by alphaXiv.