Parallel trials and subtasks cut latency in difficult LLM reasoning

The paper identifies two types of parallelism in LLM reasoning: Subtask Parallelism (decomposing tasks into independent subproblems) and Trial Parallelism (speculative exploration of competing hypotheses in parallel). Analysis of DeepSeek-V4 and DeepSeek-R1 shows Trial Parallelism accounts for the majority of parallelizable computation—65.5% on HLE for DeepSeek-V4—and grows more dominant on harder problems. The authors introduce Parason, a framework that converts sequential reasoning traces into structured parallel trajectories using a context-free grammar, then trains models with Parallelism-Aware GRPO that rewards lower latency and balanced use of both parallelism types. At inference, parallel branches are executed via tool calls, achieving an average 1.7× wall-clock speedup on AIME24 and AIME25 math benchmarks while maintaining competitive accuracy. This is interesting because it directly addresses the latency bottleneck that limits LLM reasoning in interactive applications.