Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling

Researchers introduced Collaborative Parallel Thinking, a training-free inference framework for test-time scaling in language models.

It lets parallel search branches share compact intermediate results, reducing redundant exploration during reasoning.

The method improved benchmark performance on HMMT and AIME while making inference more efficient.

Overall, it aims to improve the cost-performance balance of multi-branch LLM reasoning.