TWITTER / X
DiscoveryRSI：将递归自我改进用于科学假设生成，27B 模型产出的假设在部分指标上优于 GPT-6 与 Claude Opus 5
dair_ai 转发介绍一篇新论文 DiscoveryRSI，该工作将"递归自我改进"（RSI）应用于生命科学领域的科学假设生成。据推文称，该系统在多智能体进化搜索生成的假设树中挑选研究方向，并将其转化为可执行任务；每轮运行后会给有贡献的假设分配"信用"，回放研究流程以总结有效策略与重复失败，并将改进应用于下一轮研究。推文称，其 27B 模型生成的假设与训练截止后发表的研究相符的比例为 6.3%，高于 GPT-6 的 2.5% 和 Claude Opus 5 的 2.3%；经过两轮迭代，自动化假设质量评分从 4.86 升至 5.39，其中可证伪性提升最大。论文链接：https://arxiv.org/abs/2610.23839 。注：上述数字均为推文/作者声称，未经原文核实。
AI 解读
该帖介绍了一篇将递归自我改进（RSI）思想落到科学发现流程上的研究，属于 AI 研究类新闻，与订阅关注的 AI 模型与机器学习研究方向直接相关。值得关注的点在于：1) 方法上把 RSI 拆解为"假设树 + 多智能体进化搜索 + 信用分配 + 策略回放与改进复用"的可执行闭环，而非抽象讨论；2) 推文给出的评测显示小模型（27B）在"假设与训练截止后已发表研究相符率"这一指标上优于 GPT-6 和 Claude Opus 5，若属实，对"用 AI 加速科学假设生成"的路线是有分量的信号；3) 两轮迭代后质量评分提升主要来自可证伪性，暗示改进机制更偏向让假设更可检验，而非单纯更"新颖"。以上评测数字和效果描述均为推文作者的单方面声称，未提供论文细节核验，分析中将其作为待验证信息处理；其行业影响（如降低科研假设生成成本、启发 RSI 在其他领域的落地）属于基于推文的有限推断。原帖为转发（RT elvis），保留来源链接和 arXiv 编号即可。
推文原文
RT elvis Recursive self-improvement is one of the most discussed ideas in AI right now. I am particularly interested in how it applies to science. I saw this really cool new paper with neat ideas, especially on hypothesis generation. A good read. Specifically, they propose DiscoveryRSI, which applies RSI to scientific hypothesis generation in four life-science domains. Its 27B model produced hypotheses that match research published after its training cutoff 6.3% of the time, compared with 2.5% for GPT-6 and 2.3% for Claude Opus 5. The system grows trees of hypotheses with multi-agent evolutionary search and turns the chosen investigations into executable tasks. After each run, it assigns credit to the hypotheses that contributed, replays the research harness to find productive strategies and repeated failures, and applies accepted improvements to new investigations. Automated hypothesis quality rose from 4.86 to 5.39 after two rounds, with the largest gains in falsifiability. https://arxiv.org/abs/2610.23839 Chat with Paper: https://academy.dair.ai/papers/discoveryrsi-recursively-learning-to-discover-valuable-hypotheses-2610.23839