A new study used chess as a controlled experimental platform to systematically explore the interaction between pretraining and reinforcement learning (RL)...
A new study used chess as a controlled experimental platform to systematically explore the
interaction between pretraining and reinforcement learning (RL) in reasoning tasks. The study found that the post-training performance given the RL amount of computation can be accurately predicted by the pretraining loss, and the slope of the RL reward curve increases approximately linearly with the number of pretrained tokens. On a 1B-parameter mathematical language model, checkpoints with longer pretraining perform better and improve faster under RL.
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs ：https://info.arxiv.org/labs/index.html.
研究以国际象棋为受控实验平台，考察预训练与强化学习在推理任务中的关系。摘要称，给定强化学习计算量时，后训练性能可由预训练损失预测，奖励曲线斜率随预训练token数近似线性提升。
该材料关注推理能力从预训练延伸至强化学习后训练时的缩放规律。除国际象棋实验外，摘要还提到1B参数数学语言模型：预训练更久的检查点在强化学习阶段表现更高，提升也更快。
Aioga 判断，这项研究的价值在于尝试用预训练损失连接后训练表现，为选择进入强化学习阶段的检查点提供可检验线索。不过，现有材料未提供实验配置、误差范围及跨任务结果，结论边界仍需核对论文。
值得关注的是，如果论文中的预测关系能在更多任务和模型规模上复现，团队可能更容易评估预训练投入与强化学习收益之间的关系。但材料不足以证明该规律可直接推广到其他推理任务或更大模型。 下一步应查阅论文完整方法与实验结果，重点核验预训练损失的预测精度、奖励曲线斜率的拟合方式、国际象棋与数学任务的评测设计，以及不同计算预算和模型规模下是否保持一致。
The readable text on this page was extracted from the public source and organized with attribution, publication time and the original link. Copyright remains with the original author and publisher.
Ingestion channel: Summary aggregation · Source domain: arxiv.org
Source: HuggingFace Daily Papers（社区热门论文）
Original link: Open original source
Aioga archive: Open intelligence page
Content record: source-page · Updated: 2026-07-17T00:00:00.000Z
统一接入主流 AI 模型 API，为开发、测试与生产环境提供稳定调用入口。
立即访问 api.w173.com
Scan to open this articleAioga aggregates global AI updates and preserves source information for verification and citation.