跳转至

Parason：揭示大语言模型推理中的子任务与试验并行性

文章背景与核心概要

随着测试时推理扩展（test-time reasoning scaling）的兴起，大语言模型（LLM）解决复杂问题的能力得到了显著提升。然而，标准的自回归解码方式会顺序处理长长的推理轨迹，在处理复杂任务时往往会产生极高的延迟（甚至长达数天或数周）。虽然现有的“子任务并行”（将高阶任务分解为独立的小块）提供了一种解决方案，但它忽略了另一种至关重要的形式：试验并行（同时运行多个推测性尝试来探索、验证和聚合相互竞争的假设）。

本文介绍了 Parason 框架，旨在揭示并学习 LLM 推理中的这两种并行性。研究发现，试验并行占可并行化推理计算的绝大部分（在 HLE 中占 DeepSeek-V4 推理步骤的 65.5%），并且在更困难的问题上其主导地位愈发明显。Parason 利用上下文无关文法将顺序推理轨迹转换为结构化并行轨迹，并通过并行感知组相对策略优化（PA-GRPO）对模型进行训练，该优化器能够平衡准确性、延迟和并行度比例。在推理阶段，Parason 通过工具调用执行学习到的并行结构，从而在现实世界中实现显著的挂钟时间加速。在 AIME24 和 AIME25 等数学基准测试中，Parason 在保持竞争力的准确率的同时，实现了约 1.7 倍的平均加速比。

📌 摘要与核心亮点 (Summary)

大语言模型（LLMs）通过测试时推理扩展显著提升了问题求解能力，但标准的自回归解码仍然按顺序执行长推理轨迹，导致困难任务的延迟极高（甚至达到数天和数周）。并行推理提供了一种自然的补救方法。然而，先前的系统主要关注子任务并行，即模型学习将高阶任务分解为可以独立求解较小块。这种方法忽略了另一种普遍存在的并行形式：试验并行，即多个推测性尝试并行探索、验证和聚合相互竞争的假设。在本文中，我们推出了 Parason，它能够揭示并学习 LLM 推理中的这两种并行形式。我们的分析表明，试验并行构成了可并行化推理计算的大多数（在 DeepSeek-V4 的 HLE 推理步骤中占 65.5%），并且在难题上变得越来越占优势。在此分类法的指导下，Parason 使用上下文无关文法将顺序推理轨迹转换为结构化并行轨迹，然后使用并行感知组相对策略优化（PA-GRPO）训练模型，其奖励共同平衡了准确性、延迟和两个并行比例。在推理时，Parason 通过工具调用执行学习到的并行结构，将理论节省转化为现实世界的挂钟加速。在包括 AIME24 和 AIME25 在内的数学推理基准测试上的实验表明，Parason 实现了约 1.7 倍的平均加速，同时保持了具有竞争力的准确性。

Large Language Models (LLMs) have greatly improved problem-solving through test-time reasoning scaling. However, standard autoregressive decoding processes long reasoning traces sequentially, leading to significant latency on complex tasks (sometimes taking days or weeks). While Subtask Parallelism (decomposing high-level tasks into independent chunks) offers a solution, it misses a crucial second form: Trial Parallelism (running multiple speculative attempts to explore, verify, and aggregate competing hypotheses simultaneously).

This paper introduces Parason, a framework designed to reveal and learn both types of parallelism in LLM reasoning. Key highlights include: * Insight: Trial Parallelism accounts for the majority of parallelizable reasoning computation (65.5% of DeepSeek-V4's reasoning steps in HLE) and scales in dominance on harder problems. * Methodology: Parason uses a context-free grammar to convert sequential reasoning traces into structured parallel trajectories. It then trains models via Parallelism-Aware Group Relative Policy Optimization (PA-GRPO), optimizing a reward that balances accuracy, latency, and parallelism ratios. * Inference & Results: At inference time, Parason executes learned parallel structures through tool calls to achieve real-world wall-clock acceleration. On mathematical benchmarks like AIME24 and AIME25, Parason achieves an average speedup of ~1.7× while preserving competitive accuracy.

📥 获取论文与资源 (Access Paper & Resources)

📑 论文摘要 (Abstract)

Scaling test-time reasoning has substantially improved the problem-solving ability of large language models (LLMs), but standard autoregressive decoding still executes long reasoning traces sequentially, creating severe latency for difficult tasks (up to days and weeks). Parallel reasoning offers a natural remedy. However, prior systems primarily focus on Subtask Parallelism, where the model learns to decompose a high-level task into smaller chunks that can be solved independently. This approach overlooks another pervasive form of parallelism: Trial Parallelism, where multiple speculative attempts explore, verify, and aggregate competing hypotheses in parallel. In this paper, we introduce Parason, which reveals and learns both forms of parallelism in LLM reasoning. Our analysis identifies Trial Parallelism as the majority of parallelizable reasoning computation (65.5% in DeepSeek-V4's reasoning steps in HLE), and it becomes increasingly dominant on hard problems. Guided by this taxonomy, Parason converts sequential reasoning traces into structured parallel trajectories with a context-free grammar, then trains models with Parallelism-Aware Group Relative Policy Optimization (PA-GRPO), whose reward jointly balances accuracy, latency, and the two parallelism ratios. At inference time, Parason executes the learned parallel structure through tool calls, translating theoretical savings to real-world wall-clock acceleration. Experiments on mathematical reasoning benchmarks including AIME24 and AIME25 show that Parason achieves an average acceleration about 1.7\(\times\) while maintaining competitive accuracy.