Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning

TL;DR

The paper presents NPR, a novel framework enabling large language models to perform genuine parallel reasoning through self-distillation and reinforcement learning, achieving significant speedups and performance improvements.

Contribution

NPR introduces a self-distilled training paradigm and a parallel policy optimization algorithm for scalable, parallel reasoning in language models without external supervision.

Findings

Achieves up to 24.5% performance improvements on reasoning benchmarks.

Realizes up to 4.6x inference speedup with parallel execution.

Demonstrates 100% genuine parallel reasoning unlike autoregressive baselines.

Abstract

We introduce Native Parallel Reasoner (NPR), a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities. NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: 1) a self-distilled progressive training paradigm that transitions from ``cold-start'' format discovery to strict topological constraints without external supervision; 2) a novel Parallel-Aware Policy Optimization (PAPO) algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and 3) a robust NPR Engine that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains of up to 24.5% and inference speedups up to 4.6x. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100% genuine parallel execution, establishing a new standard for self-evolving, efficient, and scalable agentic reasoning.