DeepSeek-V4 spends 65.5% of parallel reasoning compute on trials

DeepSeek-V4 spends 65.5% of its parallelizable reasoning steps on trial parallelism, not subtask splitting, according to a paper measuring the model's reasoning traces on the HLE benchmark. Most published parallel-reasoning research targeted subtask splitting instead, the smaller share of that compute, and the paper finds the gap widens as problems get harder.

Trial parallelism runs multiple speculative attempts at once, letting a model explore, verify, and combine competing answers rather than splitting one problem into pieces.

The paper's authors built a training method called Parason to close that gap. It converts a sequential reasoning trace into a structured parallel trajectory using a context-free grammar. Training then uses a parallelism-aware variant of the GRPO reinforcement learning algorithm. On the AIME24 and AIME25 math benchmarks, Parason runs about 1.7x faster than sequential reasoning while holding accuracy roughly level.

The paper doesn't say whether the 65.5% split holds outside DeepSeek-V4 or on benchmarks other than HLE.