{"paper":{"title":"Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Large language models can learn genuine parallel reasoning on their own through self-distilled reinforcement learning.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jun Bai, Shuyi Zhang, Song-Chun Zhu, Tong Wu, Yang Liu, Yanting Wang, Zilong Zheng, Zixia Jia, Ziyong Lin","submitted_at":"2025-12-08T11:39:43Z","abstract_excerpt":"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 "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The self-distilled progressive training paradigm successfully transitions the model to native parallel cognition with strict topological constraints without external supervision or falling back to sequential behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"NPR trains LLMs to reason in parallel via self-distilled RL, delivering up to 24.5% performance gains and 4.6x speedups with 100% genuine parallel execution on reasoning benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models can learn genuine parallel reasoning on their own through self-distilled reinforcement learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"214879fc52049511adfe68c891e7a14efd86ff8859357ae9ba3aafb7195fa9a3"},"source":{"id":"2512.07461","kind":"arxiv","version":3},"verdict":{"id":"3b615a40-54df-42d0-9d2d-50cce85421b8","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T01:05:32.222004Z","strongest_claim":"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.","one_line_summary":"NPR trains LLMs to reason in parallel via self-distilled RL, delivering up to 24.5% performance gains and 4.6x speedups with 100% genuine parallel execution on reasoning benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The self-distilled progressive training paradigm successfully transitions the model to native parallel cognition with strict topological constraints without external supervision or falling back to sequential behavior.","pith_extraction_headline":"Large language models can learn genuine parallel reasoning on their own through self-distilled reinforcement learning."},"references":{"count":25,"sample":[{"doi":"","year":null,"title":"Doing: Agents that Reason by Scaling Test-Time Interaction , author=","work_id":"1619f6ef-5a19-4b4e-9731-fef04ca5be72","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Multiverse: Your language models secretly decide how to parallelize and merge generation","work_id":"292f97a6-b3ce-43b7-9296-047c8f9fecc0","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","ref_index":3,"cited_arxiv_id":"2402.03300","is_internal_anchor":true},{"doi":"","year":null,"title":"Parallel-r1: Towards parallel thinking via reinforcement learning","work_id":"75a4a861-4da2-4147-926b-d361952ab5e5","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Parallelsearch: Train your llms to decompose query and search sub-queries in parallel with reinforcement learning","work_id":"1a767f87-f215-4e70-bdcd-a05327fe1314","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":25,"snapshot_sha256":"627a65004213ae8cf3c79163fffb511675df75247d3870b24adefa23a2e6ff79","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}