Berkeley AI Research details Adaptive Parallel Reasoning for efficient inference scaling

Adaptive Parallel Reasoning allows a reasoning model to autonomously decide when to decompose and parallelize independent subtasks, how many concurrent threads to spawn, and how to coordinate them based on the problem. This addresses limitations of sequential reasoning in LLMs, where scaling linearly with exploration leads to context-rot, increased latency (tens of minutes or hours for complex tasks), and higher compute intensity. Current methods often determine parallel structure externally, but the evolution moves towards adaptive control, covering techniques like Fork-and-join, Self-consistency/Majority Voting, Tree of Thoughts, GroupThink, and Hogwild! Inference.