TokenRhythm open-sources NeoHorse-1 models, claims closed-loop step toward RSI
TokenRhythm released open-source NeoHorse-1 models in 4B and 9B sizes, trained by using a router to record agent trajectories and teaching a small model to complete tasks independently. The company describes this as the first closed-loop validation toward recursive self-improvement (RSI), with infrastructure support from Infinigence and research collaboration from Tsinghua and Peking University teams.
Coverage timeline
机器之心机器之心
编辑 | 泽南 Agent 完成任务后，学到的经验去了哪里？ 一个 Agent 接到项目排期任务，它找到了工作目录里的文件，大致理解用户想要什么，却漏读了一封包含最新消息修改的邮件。随后，它沿着过时信息继续规划，生成了无效排期，最后还把文件写到了错误位置。 执行链在中途断掉，随后就全是基于错误推理的错误答案。看到之后，你显然会让 AI「重新审视问题」再生成一遍。问题在于，下一次怎么办？ 今天的大多数系统中，Agent 的经验都留在日志里，模型每天在真实环境里产生的执行经验，几乎从不进入模型参数。这意味着模型下一次遇到相似任务，依旧可能踩一遍同样的坑。 基元律动最近发布的 NeoHorse-1 试图改变这件事。他们开源了 4B 与 9B 两个尺寸的模型，训练出来的思路听起来有点绕：让一个本来负责给模型「派活」的路由系统，顺便把所有模型走过的路记录下来，再把这些路教给一个小模型，让它自己就能走完一群模型走过的路。 论文链接：https://arxiv.org/abs/2609.08183 HuggingFace：https://huggingface.co/collections/TokenRhythm/neohorse-1 GitHub：https://github.com/TokenRhythm/NeoHorse 这件事， 被他们称为迈向 RSI 的第一次闭环验证 。 该模型由无问芯穹提供底层基础设施支撑与 Infra 优化技术能力，清华大学、北京大学团队参与算法和训练方法研究，共同探索提升 Agent 后训练的数据利用效率和训练效果。 大模型的终极目标：自我迭代 RSI（Recursive Self-Improvement，递归自我改进）是 2026 年 AI 领域一个炙手可热，也富有争议的词汇。它的愿景并不复杂：让 AI 把自己的产出作为经验重新喂回研发流程，参与设计更强的下一代自己，循环往复、自我升级。 今年 6 月，Anthropic 发布《When AI builds itself》，称在内部观察到了 RSI 的早期迹象。这个星期，OpenAI 发布博客首次公开了自己 RSI 的进展。OpenAI 称，其已经实现去年秋天定下的目标：在今年 9 月前造出一个「自动化研究实习生」。 但 Anthropic、OpenAI 自己也承认完整的 RSI 时代尚未到来，
Related stories
Hinton co-authors first paper on recursive self-improvement and intelligence explosion
Geoffrey Hinton, Yoshua Bengio, Andrew Barto, Jakub Pachocki, and 18 other researchers co-authored a paper titled "What if automating AI R&D triggers an intelligence explosion?" The paper examines whether automating AI research and development could close the loop of recursive self-improvement, potentially compressing years of AI progress into months or even about five weeks. It argues that an intelligence explosion might not require superintelligence first, but could start if the automation efficiency of AI R&D is high enough.
1 reports · 1 sources
ModularRSI: Frozen-Weight Agents Improve via Harness Self-Modification
IQuest Research, with Beihang University and the University of Manchester, released ModularRSI, a method for agent continuous self-improvement that modifies the execution harness while keeping the base model frozen. On Terminal-Bench 2.0, accuracy rose from 47.57 to 52.43 through harness evolution, with improvements transferring to unseen tasks, domains, and base models. The work defines an agent as A=(M,H), where H is the harness controlling environment observation, context maintenance, tool invocation, failure recovery, and task termination.
1 reports · 1 sources
Agibot Psi-R2.5 Embodied Foundation Model Uses Paired Human-Robot Data for One-Shot Imitation
Agibot (Lingchu Intelligence) released Psi-R2.5, a new embodied foundation model that introduces a training pipeline based on paired human-robot data (Pair Data) instead of the raw large-scale human videos used previously. After a single human demonstration, the model can reproduce a task through context learning without retraining or fine-tuning. The new approach aims to streamline the conversion of human data into robot-executable trajectories, addressing data scarcity and embodiment gap issues.
Recuris: Recursive Memory Self-Improvement Boosts LLM Agents from 3B to Claude Opus 5
Researchers from NUS, Princeton, Stanford, and Oxford introduced Recuris, the first agent architecture applying recursive self-improvement to memory control. It addresses memory retrieval and iteration failures in long-horizon tasks via three core techniques, achieving training-free performance gains across models from 3B to Claude Opus 5. The paper and code are publicly available.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe