Atria Dawn: The Dawn of Agentic Superintelligence

🤗 Upvotes: 159 | cs.AI

Authors:
Honglin Guo, Tao Gui, Yicheng Chen, Guanting Dong, Qiming Ge, Yuyang Hu, Zixian Huang, Jiajie Jin, Alexander Lam, Yining Li, Jiahang Lin, Yanjiang Liu, Xinyu Lu, Haijun Lv, Junlin Shang, Qisheng Su, Guoqiang Wang, Rui Wang, Zhecan Wang, Hao Xiang, Xinchen Xie, Shuhao Xing, Xiaoyu Xing, Wanghan Xu, Xinyu Yang, Yajie Yang, Chengfeng Zhao, Haoran Zhao, Ruojun Zhou, Yunhua Zhou, Yicheng Zou, Kun Cai, Qiye Cai, Xinmeng Che, Haodong Chen, Jiabei Chen, Jiahao Chen, Jiayi Chen, Yujia Chen, Lizhi Cui, Youheng Dai, Xin Deng, Yi Dong, Shihan Dou, Chenya Gu, Xu Guo, Ding Han, Feiyang Hao, Haotan He, Jie Hou, Binze Hu, Zijian Hu, Junhao Huang, Huicheng Jiang, Jiazhen Jiang, Shufan Jiang, Jiahao Kuang, Bowen Lai, Bo Li, Jiaqiang Li, Peng Li, Qilong Li, Zhuoqun Li, Jiaxiang Liu, Shuainan Liu, Tong Liu, Yi Liu, Zhonghang Lu, Jianwen Luo, Yanyi Luo, Huijie Lv, Ningsheng Ma, Zerun Ma, Houcheng Min, Chengjun Pan, Qiyuan Peng, Xiaoxuan Peng, Jianmin Qian, Jiantao Qiu, Wanying Ren, Huayu Sha, Jifei Shan, Zixin Shang, Bing Shao, Zhuohui Sheng, Jiayang Shi, Yang Shu, Aierpanjiang Simayi, Sirui Song, Yuxiao Song, Zhe Sun, Zhichao Sun, Wenzhe Tan, Wenhui Tian, Zhongbo Tian, Hanchen Wang, Pengbo Wang, Rui Wang, Yiding Wang, Yuhui Wang, Zhiheng Xi, Caijun Xu, Chao Xu, Yongfeng Xu, Xiaolei Yang, Zhixiong Yang, Qian Yao, Shihong Yi, Yuankai Ying, Jia Yu, Dingbo Yuan, Hao Yuan, Junjie Yuan, Bo Zhang, Caixian Zhang, Qiuyinzhe Zhang, Jiyuan Zhao, Penghao Zhao, Ying Zhao, Pujun Zheng, Xiaoxue Zhong, Xiaohao Zhou, Xinyu Zhou, Dongsheng Zhu, Guanru Zhu, Yulun Zhu, Yaojie Lu, Tao Ji, Hongyu Lin, Yutao Zhu, Pengfei Cao, Guoxiu He, Xianpei Han, Ben He, Zhicheng Dou, Kang Liu, Qi Zhang, Le Sun, Jun Zhao, Ji-Rong Wen, Xuanjing Huang, Yu-Gang Jiang, Bowen Zhou, et al. (43 additional authors not shown)

Title:
Atria Dawn: The Dawn of Agentic Superintelligence

Arxiv:
http://arxiv.org/abs/2609.15818v1

Abstract:
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.

What is Daily Paper Cast?

We update every weekday to discuss highest-voted papers from Huggingface Daily Paper (https://huggingface.co/papers). Both the podcast scripts and audio are generated by AI. Feedback and suggestions are welcome! Email us: [e-mail]

Creator:
Jingwen Liang, 3D ML, https://www.linkedin.com/in/jingwen-liang/
Gengyu Wang, LLM ML, http://wanggengyu.com

Listen on:
Spotify: https://open.spotify.com/show/21nrhmdaA8qoBiH8q03NXL
Apple Podcast: https://podcasts.apple.com/us/podcast/daily-paper-cast/id1777620236

Cover Image by Kawen Kuang https://kawen.art

Evan: Welcome to Daily Paper Cast.

Evan: Today’s paper is from the Hugging Face daily paper list, dated September 15, 2026, and it's garnered 159 upvotes.

Ashley: The paper is titled 'Atria Dawn: The Dawn of Agentic Superintelligence.'

Ashley: The first two authors are Honglin Guo and Tao Gui, with Tao Gui being the corresponding author from Fudan University.

Evan: So, what does the Introduction tell us about this paper?

Ashley: To begin, the Introduction sets the stage by discussing the role of language-model agents in tasks that require sustained tool use, such as software engineering and workplace document reasoning.

Evan: Right, and they mention that these agents are increasingly participating in the development of AI systems themselves.

How do they contribute?

Ashley: Exactly.

The paper highlights that agents can propose candidates, run experiments, and revise solutions in response to feedback.

This is supported by studies on algorithm discovery and automated research.

Evan: That sounds impressive.

But what key questions are they aiming to answer through this research?

Ashley: A significant question they aim to tackle is around the division of responsibility in a real model-development project.

Specifically, who identifies worthwhile problems, chooses methods, interprets results, and decides the next steps?

Evan: Got it.

And to address these questions, they introduce Atria Dawn Preview, right?

Ashley: Yes.

Atria Dawn Preview is a foundation agentic language model specifically designed for scientific research and engineering workflows.

It's built on a 744-billion-parameter mixture-of-experts model.

Evan: That’s substantial.

What’s unique about its training process?

Ashley: The model is trained using a Verifiable Experience Pipeline, which links tool-mediated interactions to executable environments and externally verified outcomes.

This training method aims to teach the model behaviors that generalize well across tasks.

Evan: Interesting.

How do the results pan out compared to other models?

Ashley: Atria Dawn Preview was evaluated on 16 benchmarks covering research, digital work, software engineering, and more.

It achieved the highest score on five of these benchmarks and was competitive across the others.

Evan: Excellent.

And what about practical applications?

Any highlighted examples?

Ashley: To provide practical insight, the paper uses selected cases that show how the model uses tools, responds to feedback, and produces artifacts.

They also analyzed the development process as a case study of human–AI collaboration.

Evan: That's a valuable perspective.

How did they conduct this analysis of human–AI collaboration?

Ashley: The analysis is based on 769 tasks logged from 56 participants, alongside agent logs.

They observed that one-third of the AI-assisted tasks were rated infeasible under the same scope and resource constraints without AI.

Evan: So, it's not just about task execution but also about how agents and humans collaborate on these tasks.

Ashley: Precisely.

Tasks included aspects like proposing methods, implementing revisions, and humans primarily handled final decisions and guided experiments through judgment and feedback.

Evan: So, the Introduction outlines the evolving role of AI agents from task execution to project-level partnership and emphasizes the need for human oversight.

Ashley: Yes, it's critical that human researchers concentrate on defining meaningful directions and ensuring accountable human authority over AI development.

Evan: That's the end of the Introduction section.

Evan: Now, let's move on to the core of the paper—how Atria Dawn Preview is built and evaluated.

Ashley, can you walk us through the training process?

Ashley: Sure, Evan.

The training of Atria Dawn Preview hinges on something called the Verifiable Experience Pipeline.

Essentially, this pipeline ensures that each training task is linked to a real execution environment where the model interacts with tools, observes states, and responds to feedback, resulting in externally verified outcomes.

Evan: Externally verified outcomes?

How does that play into training the model?

Ashley: Great question.

The outcomes are verified using a range of domain-specific signals.

These could be tests, metrics, file state checks, geometric structures, or even human-defined criteria.

Only those experiences that include a task, its trajectory, artifacts, and verification evidence are incorporated into the model’s capabilities.

Evan: So, this pipeline doesn’t just generate data; it verifies its relevance and correctness.

What happens during a typical training rollout?

Ashley: In a rollout, the agent observes the environment, makes decisions on which tools to use, inspects their outputs, and adapts its actions based on feedback.

For example, if an experiment metric doesn’t meet a threshold, the agent can modify its approach and try again.

Evan: It sounds like an intricate way to simulate a real-world research environment.

Ashley: Exactly.

One interesting aspect is failure analysis.

By analyzing recurring problems such as ineffective tool selection or incomplete verification, the process can guide future task construction and environment refinement.

Even failed runs are valuable as they can serve as diagnostic cases.

Evan: Does the paper discuss how they evaluate Atria Dawn Preview’s performance?

Ashley: They do.

Evaluation is conducted across 16 benchmarks covering tool use, search and research, professional tasks, software engineering, ML engineering, and cybersecurity.

This suite of benchmarks provides a comprehensive view of the model’s capabilities in different domains.

Evan: Can you share some of the specific benchmarks mentioned?

Ashley: Of course.

For example, on AutomationBench, which evaluates automation tasks, Atria Dawn achieved the highest reported score of 53.8, surpassing other models like Qwen 3.8 Max and GLM 5.3.

It also excelled on BFCL v4, scoring 77.0, a notable 2.9 points over the runner-up.

Evan: That’s impressive.

How does it perform in software engineering tasks?

Ashley: In the domain of software engineering, Atria Dawn was tested on SWE-bench Pro.

Here, it achieved a HumanRank score of 59.6, competing closely with other models.

Additionally, in the cybersecurity focus through CyberGym, it scored 86.5, which indicates its strength in identifying and addressing security vulnerabilities.

Evan: So, we’ve got a model that’s high-performing across a vast span of tasks.

How about insights from concrete case studies?

Any specific examples?

Ashley: Yes, the paper details some illustrative cases.

One example is a scientific research task involving weather forecasting.

Atria Dawn processed over 100 gigabytes of weather data and built a vision-transformer-based prediction model.

They also have a case on improving the performance of a Gated Delta Network through iterative experimentation and adjustment.

Evan: Those are high-complexity tasks.

What about cases beyond research, say, in software or applications?

Ashley: In a case involving software development, Atria Dawn successfully built a Minimal Operating System, or MiniOS, complete with a persistent file system and autorun capabilities across two QEMU sessions.

They also demonstrated practical applications like CAD assemblies and narrative games.

Evan: Oh, wow!

This really showcases its versatility.

What else did they evaluate?

Ashley: The paper also covers deliverables like professional reports and presentations.

Atria Dawn created detailed reports on topics like clean energy siting and healthcare investment, incorporating evidence, calculations, tables, and charts.

Evan: So, in essence, it’s not just automating simple tasks but generating valuable professional artifacts

Ashley: Exactly.

The final set of examples comes from cybersecurity.

Atria Dawn was able to perform end-to-end tasks, from diagnosing security vulnerabilities to implementing and verifying fixes within an authorized test environment.

Evan: Quite a range of capabilities indeed.

What does the paper say about the broader implications of this development?

Ashley: The broader implication is the shift in AI’s role from executing individual tasks to becoming a project-level partner.

This involves generating methods, implementing revisions, and even planning research endeavors while relying on humans for final judgment and guidance.

Evan: That’s profound.

It really highlights the evolving nature of human–AI collaboration.

Ashley: Indeed, and it emphasizes the importance of maintaining meaningful human oversight even as AI takes on more complex roles.

Evan: And with that, we wrap up the Method section.

Evan: Let's dive into the Experiment and Results section of the paper.

Ashley, how did they evaluate the collaborative dynamics between humans and AI?

Ashley: The researchers conducted an extensive analysis of the development process of Atria Dawn, using 769 logged tasks from 56 participants.

They focused on understanding how responsibilities were distributed and how effective the collaboration was.

Evan: What key metrics or approaches did they use for the evaluation?

Ashley: They looked at three main aspects: the feasibility of tasks with and without AI, the distribution of roles in method proposal and selection, and how revisions and feedback cycles operated between humans and the agents.

Evan: Interesting.

Let’s start with task feasibility.

What did they find?

Ashley: Among 455 completed AI-assisted tasks, 33.2% were rated as infeasible without AI.

This highlights the significant role AI played in expanding what could be achieved within the project’s scope and resources.

Evan: That’s substantial.

What does the data show about how roles were distributed between the humans and AI during method proposal and selection?

Ashley: In terms of proposing methods and parameters, AI was often the initiator, proposing 55.4% of the options, but humans made the final decisions 85.5% of the time.

When it came to goals or scope, humans were even more involved, making the final decisions in 93.4% of the cases.

Evan: So, AI mainly generates options, and humans choose among them.

How about when things don’t go as planned?

Ashley: When difficulties arose, human intervention was critical.

About 76% of tasks moved forward due to some form of human intervention, whether that was adding context, diagnosing the issue, or changing the method.

Evan: Can you give a bit more detail on how human feedback impacted the subsequent steps?

Ashley: Certainly.

Out of 627 tasks where revisions were needed, 75.4% of the substantive revisions were performed by the AI after receiving human feedback.

So, human oversight was crucial in guiding the agent on what needed change.

Evan: It seems like a continuous feedback loop.

Do they provide any illustrations or charts to support their findings?

Ashley: Yes, they include several figures to visualize their data.

For example, Figure 6 on page 9 shows the rising ratio of agent actions per human prompt over time, revealing increased reliance on AI with human oversight growing more targeted.

Evan: And what about the specific nature of the tasks?

Did they examine different categories of work?

Ashley: They did.

In their analysis, participants estimated the time needed to complete tasks without AI across various categories.

For example, tasks like tool harnesses and infrastructure saw a higher rate of infeasibility without AI, further stressing AI’s crucial role in these areas.

Evan: That's a lot of data to unpack.

What’s the overarching takeaway from these results?

Ashley: The main takeaway is that AI agents, while highly capable, still depend heavily on human judgment for steering research direction and making final decisions.

It’s a blend where AI handles execution but humans provide the critical oversight and context.

Evan: So, the experiment and results show a dynamic interaction where AI significantly amplifies human effort, but humans remain the ultimate decision-makers.

It’s a partnership rather than full autonomy.

Ashley: Exactly.

And it calls for continued emphasis on human researchers to guide, evaluate, and make final decisions, ensuring both progress and accountability in AI-assisted research and development.

Evan: That's the end of the Experiment section.

Evan: Let's move on to the Related Work section of the paper.

How does Atria Dawn Preview fit within the broader landscape of AI research, Ashley?

Ashley: The Related Work section serves as a contextual framework, linking Atria Dawn Preview's contributions to existing research in the field of AI-driven collaboration and agentic models.

The paper references a series of studies that surround the evolution and application of AI agents in complex tasks.

Evan: What foundational studies does the paper build upon?

Ashley: The paper builds upon foundational studies like the work by Schick et al.

in 2023 on tool-mediated reasoning, which was crucial in forming the concept of agents using tools to perform tasks effectively.

Also, it references Shen et al.'s studies in 2026 on sustained tool use in software engineering, providing a key underpinning for understanding the dynamics of these tasks.

Evan: It's good to see such a solid foundation.

How does it tie into recent advancements?

Ashley: It ties into recent advancements by discussing how AI systems increasingly contribute to developing their successors.

For example, Jimenez et al., 2024, focused on software engineering tasks, while Yang et al., 2026a, explored workplace document reasoning.

These aspects are integrated into the development process of Atria Dawn Preview.

Evan: How do the studies on algorithm discovery and automated research play a role here?

Ashley: Studies on algorithm discovery and automated research, such as the work by AlphaEvolve Team in 2025 and Lu et al.

in 2026, demonstrated how AI agents could propose candidates, run experiments, and revise solutions autonomously.

This is essential for understanding Atria Dawn’s role in identifying methods and running iterative improvements, pushing the boundaries of agentic models.

Evan: Certainly.

And how about the aspect of human judgment in AI collaboration?

Ashley: Human judgment remains a cornerstone of AI collaboration.

The paper references Guo et al., 2026, who examined workplace document reasoning and highlighted how human researchers' feedback and judgment guide AI's exploratory efforts.

Human oversight ensures the reliability and direction of AI-driven research endeavors.

Evan: What are the key takeaways from these past works regarding human–AI collaboration?

Ashley: A key takeaway is the evolving role of AI agents from merely executing tasks to acting as collaborators.

Novikov et al., 2025, and Xi et al., 2025, pointed out how AI agents increasingly take on project-level roles, proposing methods and iterating on research ideas, while human researchers focus more on high-level judgment and decision-making.

Evan: It sounds like we’re seeing a real partnership evolving here.

What do they say about the advancements in AI recursive self-improvement?

Ashley: For recursive self-improvement, studies by OpenAI and Anthropic in 2026 are particularly relevant.

They discussed how stronger models could contribute more effectively to R&D, creating self-reinforcing cycles of capability gains.

This is echoed in Atria Dawn’s effort to ensure sustained improvement through its training pipeline and iterative task construction.

Evan: How do other papers view the balance between AI autonomy and human intervention?

Ashley: Other papers like those by Burns et al., 2024 and Chen et al., 2021, highlight the balance between AI's ability to automate tasks and the necessity of human steering.

This balance is critical in ensuring that AI models evolve in a controlled and predictable manner, without compromising oversight.

Evan: Do they mention frameworks or systems that have inspired Atria Dawn’s development?

Ashley: Yes, the Verifiable Experience Pipeline used in Atria Dawn is inspired by several frameworks.

For instance, Patil et al., 2025, discussed tool-mediated reasoning for agentic evaluation.

Feng et al., 2026a, introduced test-time training approaches that help AI adapt and learn during actual task execution, influencing Atria Dawn’s dynamic training process.

Evan: It seems like a synthesis of multiple pioneering efforts.

How does this blend of past research inform future directions?

Ashley: The blend of past research informs future directions by emphasizing the need for systems that can autonomously handle R&D tasks while maintaining human judgment.

As the paper notes, continuous advancements in AI capabilities must be matched with robust oversight and alignment protocols to ensure safety and accountability.

Evan: Indeed, the related work underscores the importance of leveraging prior advancements while keeping humans in the loop.

Ashley: By synthesizing these references, Atria Dawn Preview not only pushes the frontier of AI research capabilities but also poses crucial questions about the evolving role of human oversight in these processes.

Evan: That’s the end of the Related Work section.

Evan: We've covered a lot of ground today.

Let's wrap up by summarizing the key contributions and takeaways from the paper on Atria Dawn: The Dawn of Agentic Superintelligence.

Ashley: Sure, Evan.

The paper introduces Atria Dawn Preview, an agentic language model designed for scientific research and engineering tasks.

It’s built on a massive 744-billion-parameter mixture-of-experts model and trained using a unique Verifiable Experience Pipeline.

Evan: And this Verifiable Experience Pipeline is what sets it apart, right?

Ashley: Exactly.

This pipeline ensures that each training task is connected to real execution environments and validated through domain-specific signals, adding a layer of reliability and relevance to the model’s capabilities.

Evan: On the performance side, Atria Dawn Preview was evaluated across 16 benchmarks and achieved the highest scores on five of them.

Ashley: Yes, from automation tasks on AutomationBench to cybersecurity challenges assessed by CyberGym, Atria Dawn demonstrated competitive performance across various domains, showcasing its versatility and robustness.

Evan: The case studies were particularly intriguing.

From processing huge datasets for weather forecasting to building a Minimal Operating System, the practical applications were impressive.

Ashley: The paper also underscores the evolving dynamics of human–AI collaboration.

While AI agents are taking on more complex roles, human researchers remain crucial for oversight, judgment, and guiding research directions.

Evan: That balance between AI autonomy and human oversight is critical to ensuring that AI development remains safe and productive.

Ashley: Indeed, and it poses important questions about how we structure these collaborations moving forward to maximize both efficiency and accountability.

Evan: And with that, we conclude today's deep dive into Atria Dawn: The Dawn of Agentic Superintelligence.

We hope you found the discussion insightful.

Ashley: Be sure to join us next time on Daily Paper Cast for more discussions on cutting-edge AI research.

If you enjoyed this episode, don't forget to subscribe and leave us a review.

Evan: Thanks for listening, everyone.

Until next time!