我们研究如何通过两种互补方法扩展竞赛编程中的推理令牌预算：训练时的强化学习（RL）和测试时的并行思维。在RL训练过程中，我们观察到验证准确率与连续检查点中生成推理令牌的平均数之间存在近似对数线性关系，并展示了两种方法来调整此训练轨迹：验证RL预热提高起始点，而随机截断在观察范围内产生更陡峭的趋势。由于在完全注意力下扩展单次生成推理令牌很快变得昂贵，我们引入了一种多轮并行思维流程，将令牌预算分布在多个线程和生成、验证和细化的轮次中。我们在此流程上端到端训练模型，使训练目标与测试时的结构相匹配。从Seed-OSS-36B开始，具有16个线程和每个线程16轮的完整系统，平均使用760万个令牌 per problem，在pass@16上匹配底层RL模型的oracle性能，并在AetherCode的456个困难竞赛编程问题上超越了GPT-5-high。
英文摘要：
We study how to scale reasoning token budgets for competitive programming through two complementary approaches: training-time reinforcement learning (RL) and test-time parallel thinking. During RL training, we observe an approximately log-linear relationship between validation accuracy and the average number of generated reasoning tokens over successive checkpoints, and show two ways to shift this training trajectory: verification RL warmup raises the starting point, while randomized clipping produces a steeper trend in the observed regime. As scaling single-generation reasoning during RL quickly becomes expensive under full attention, we introduce a multi-round parallel thinking pipeline that distributes the token budget across threads and rounds of generation, verification, and refinement. We train the model end-to-end on this pipeline to match the training objective to the test-time structure. Starting from Seed-OSS-36B, the full system with 16 threads and 16 rounds per thread matches the underlying RL model's oracle pass@16 at pass@1 using 7.6 million tokens per problem on average, and surpasses GPT-5-high on 456 hard competitive programming problems from AetherCode.
The Price of a Second Thought: On the Evaluation of Reasoning Efficiency in Large Language Models
University of Electronic Science and Technology of China（电子科技大学）；Beijing Academy of Artificial Intelligence（北京人工智能研究院）；Nanyang Technological University（南洋理工大学）
SwingArena: Competitive Programming Arena for Long-context GitHub Issue Solving
The University of Hong Kong（香港大学）；University of California Los Angeles（加州大学洛杉矶分校）；Tsinghua University（清华大学）；University of Michigan Ann Arbor（密歇根大学安娜堡分校）；The Ohio State University（俄亥俄州立大学）；University of Edinburgh（爱丁堡大学）；The Chinese University of Hong Kong（香港中文大学）；The Hong Kong Polytechnic University（香港理工大学）；Hong Kong University of Science and Technology (Guangzhou)（香港科学与技术大学（广州））；LMSYS Org（LMSYS组织）
InteractBench: Benchmarking LLMs on Competitive Programming under Unrevealed Information
School of Cyber Science and Engineering, Huazhong University of Science and Technology（华中科技大学网络空间科学与工程学院）；Key Laboratory of Cyberspace Security, Ministry of Education（教育部网络空间安全重点实验室）；Hubei Key Laboratory of Distributed System Security（湖北省分布式系统安全重点实验室）
AutoCode: LLMs as Problem Setters for Competitive Programming
University of California San Diego（加州大学圣地亚哥分校）；New York University（纽约大学）；University of Washington（华盛顿大学）；Princeton University（普林斯顿大学）；University of California Berkeley（加州大学伯克利分校）；OpenAI；Massachusetts Institute of Technology（麻省理工学院）；University of Waterloo（滑铁卢大学）；Sentient Labs
ReMA: Learning to Meta-think for LLMs with Multi-Agent Reinforcement Learning
Shanghai Jiao Tong University（上海交通大学）；Shanghai Artificial Intelligence Laboratory（上海人工智能实验室）；University of British Columbia（不列颠哥伦比亚大学）；University College London（伦敦大学学院）
PRIME: Planning and Retrieval-Integrated Memory for Enhanced Reasoning
Center for Healthcare Organization and Implementation Research, VA Bedford Health Care（VA Bedford Health Care 医疗组织与实施研究中心）；Manning College of Information and Computer Sciences, University of Massachusetts Amherst（马萨诸塞大学阿默斯特分校信息与计算机科学学院）；Miner School of Computer and Information Sciences, University of Massachusetts Lowell（马萨诸塞大学洛厄尔分校 Miner 信息与计算机科学学院）
LiveCodeBench Pro: How Do Olympiad Medalists Judge LLMs in Competitive Programming?
New York University（纽约大学）；Princeton University（普林斯顿大学）；University of California San Diego（加利福尼亚大学圣地亚哥分校）；University of Washington（华盛顿大学）；Canyon Crest Academy（峡谷克雷斯特学院）；University of Waterloo（滑铁卢大学）；McGill University（麦吉尔大学）
