Atria Dawn: The Dawn of Agentic Superintelligence (Technical Report)

Abstract (opencli arxiv paper / abs-jina fallback)

arXiv:2609.15818, cs.AI, v1 2026-09-14 16:22 UTC, 23 pages, 10 figures, comments link github.com/atria-asi/Atria-Dawn-Preview. 43+ authors incl. Honglin Guo, Tao Gui, Yicheng Chen, Guanting Dong, Qiming Ge et al. Abstract (jina-fetched): 'As AI agents become participants in the development of their successors, they reshape both the production of intelligence and the role of human researchers. We introduce Atria Dawn Preview, a foundation agentic language model designed for scientific research and engineering workflows, with the goal of expanding the frontier of agent productivity in the real world. This model is trained via a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments and externally verified outcomes. Across 16 benchmarks spanning real-world research, engineering, and digital work, Atria Dawn Preview is competitive with frontier agents and achieves the highest reported score on five of them. Beyond standalone performance, we examine the real research-and-development process behind this model as a case study of human–AI collaboration, analyzing 769 task records from 56 participants together with agent logs. When asked to evaluate completed tasks under comparable conditions, participants rated about one-third of completed AI-assisted tasks as infeasible without AI. More strikingly, agents frequently propose methods and implement revisions, while humans retain most final decisions and guide exploration through judgment and feedback. These observations indicate a shift from task-level execution to project-level partnership, with human effort concentrating on what is worth pursuing and how evidence should guide research. Progress toward more autonomous AI research must therefore advance both the capacity for discovery and the capacity for meaningful human oversight, preserving accountable human authority over the risks and direction of continued development.'

Note: opencli arxiv paper 2609.15818 returned HTTP 429 on 2026-09-15; this entry is grounded from the abs-page metadata captured earlier the same day.

论文要点 (中文)

Shanghai AI Lab（Atria Team，含 Honglin Guo / Tao Gui / Yicheng Chen / Guanting Dong / Qiming Ge 等 43+ 作者）2026-09-14 在 arXiv 公开 Atria Dawn Preview 技术报告（cs.AI，v1 16:22 UTC，23 页 10 图）。论文把 AI agent 在「后继模型的研发中既是执行者也是参与者」作为前提，重新审视智力的生产与人类研究者的角色；模型通过 Verifiable Experience Pipeline（VEP）训练，把工具调用与可执行环境、外部可验证结果相连。在 16 个跨真实研究、工程、数字工作的基准上，与 frontier agent 相当并在 5 个上达到目前已报告的最高分。论文同时给出研发流程的实证：分析 56 位参与者、769 条任务记录与 agent 日志；约三分之一完成的任务被评为「无 AI 不可行」；agent 经常提出方法并实现修订，但人类保留最终决策、通过判断与反馈引导探索方向。结论是从「任务级执行」向「项目级伙伴关系」的转变，把 AI 研究进展同时推到发现能力与有意义的人类监督能力两侧，保留可问责的人类权威。开源仓 github.com/atria-asi/Atria-Dawn-Preview；HF 权重 internlm/Atria-Dawn-Preview（MIT，256K 上下文，744B MoE GLM-5.2 基座——卡表数字以官方为准，未独立复测）。

Key claims (English)

Atria Dawn Preview (Shanghai AI Lab; arXiv:2609.15818; cs.AI; v1 2026-09-14 16:22 UTC; 23 pages, 10 figures; 43+ authors) frames AI agents as participants in their successors' development, redefining the production of intelligence and the role of human researchers. Training is through a Verifiable Experience Pipeline that ties tool-mediated interactions to executable environments and externally verified outcomes. Across 16 benchmarks spanning real research, engineering, and digital work, the model is competitive with frontier agents and reaches the highest reported score on five. As a human–AI collaboration case study, the paper analyzes 769 task records from 56 participants alongside agent logs: about one-third of completed AI-assisted tasks are rated infeasible without AI; agents frequently propose methods and implement revisions while humans retain most final decisions and steer exploration through judgment and feedback. The takeaway is a shift from task-level execution to project-level partnership, advancing both the capacity for discovery and the capacity for meaningful human oversight, while preserving accountable human authority. Open-source repo github.com/atria-asi/Atria-Dawn-Preview; HF weights internlm/Atria-Dawn-Preview (MIT, 256K context, 744B MoE built on GLM-5.2 — model-card numbers are the author's harness, not independent replications).

Obsidian 证据摘录

「Atria Dawn Preview 把长程 agent 权重开出来」+「744B 参数 MoE GLM-5.2；上下文 256K；许可 MIT」+「模型卡 AutomationBench 53.8、BrowseComp 92.5、CyberGym 86.5——这是官方 harness 数字，不是第三方复测」——Hermes 定时任务/X-每日简报/2026-09-14-X-Hot-Brief_AtriaDawn-OM1-GlassImaging.md L38-39

链接

论文：https://arxiv.org/abs/2609.15818
HF 权重：https://huggingface.co/internlm/Atria-Dawn-Preview
GitHub：https://github.com/atria-asi/Atria-Dawn-Preview
官方站点：https://atria-asi.ai/
DeepResearch abs 抓取：DeepResearch/2026-09-15-evening-AtriaDawn-Preview-研究材料/references/04-arxiv-abs-jina.md
X 简报：Hermes 定时任务/X-每日简报/2026-09-14-X-Hot-Brief_AtriaDawn-OM1-GlassImaging.md