PaperlayerView exact revision on arXiv

Source-linked AI summary

Atria Dawn: The Dawn of Agentic Superintelligence

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

arXiv:2609.15818v1cs.AI

TL;DR

Task scores alone do not reveal how humans and agents divide research judgment in real model development. This paper introduces Atria Dawn Preview and analyzes its development, reporting competitive results across 16 benchmarks and agents proposing methods while humans retain most final choices.

  • Problem

    Task scores alone do not show who identifies worthwhile problems, chooses methods, interprets uncertain results, and decides during real model development.

    Source: §1 · p. 2

  • Method

    The paper trains Atria Dawn Preview with a Verifiable Experience Pipeline and studies its development using 769 task records and logs from 56 participants.

    Source: §1 · p. 2, §6 · p. 14

  • Results

    Across 16 benchmarks, Atria Dawn Preview is competitive with frontier agents and achieves the highest reported score on five; agents proposed 64.6% of methods and decisions, while humans made the final choice in 85.5%.

    Source: §1 · p. 2, §6 · p. 14

  • Takeaways & Limitations

    The paper argues that AI capability advances should be matched by alignment and oversight investments, with clear accountability for alignment objectives and associated risks.

    Source: §5.2 · p. 14

    Agent capabilities do not ensure reliable selection of research tasks or strategies, and interpreting uncertain results and choosing directions requires research judgment.

    Source: §3.2 · p. 8

Abstract

from arXiv · show

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.

1. Introduction

Atria Dawn Preview pairs a tool-trained agentic model with a case study of its development, finding strong but domain-variable benchmark performance and a human–AI division of labor in which agents propose and implement while humans evaluate and steer.

1. Introduction

  • The introduction frames agent participation in AI development as raising questions about who identifies worthwhile problems, selects methods, interprets uncertain results, and decides what to pursue.It argues that recursive self-improvement requires strengthening the research process, while human judgment and oversight remain consequential.Source: §1 · p. 2
  • Atria Dawn Preview achieves the highest score on five of 16 benchmarks, with performance varying across research, digital work, software engineering, and other agentic domains.The paper also presents selected cases of tool use, feedback response, and artifact production.Source: §1 · p. 2
  • Built on a 744-billion-parameter mixture-of-experts foundation model, Atria Dawn Preview uses a Verifiable Experience Pipeline linking agent tasks and trajectories to artifacts and externally checked outcomes.It is designed for scientific research and engineering workflows.Source: §1 · p. 2
  • In a case study of Atria Dawn’s development, participants rated roughly one-third of completed AI-assisted tasks infeasible without AI under comparable scope and resource constraints.Analysis of 769 task records from 56 participants and agent logs found agents often initiated approaches and executed changes, while humans evaluated, selected, and steered inquiry.Source: §1 · p. 2

2. Atria Dawn Preview

Atria Dawn Preview is an agentic model for research and engineering, trained through tool-mediated experience tied to externally verified outcomes. It leads or remains competitive across diverse benchmarks and demonstrates workflows in scientific research, software creation, professional delivery, and cybersecurity.

2.1. Training Overview

  • The Verifiable Experience Pipeline trains reusable behavior by connecting tool-using trajectories and artifacts to external outcome checks, then curating examples and using failure analysis to refine tasks.Rollouts adapt to feedback; verification can use tests, metrics, file or application state, geometric checks, source evidence, or human-defined criteria.Source: §2.1 · p. 3

2.2. Evaluation Summary

  • Across 16 benchmarks, Atria Dawn has the highest reported score on five and the second-highest on three, with frontier capability spanning tool use, search, and workspace execution.Rankings use available entries; missing results are not treated as zero.Source: §2.2 · pp. 3–4, Table 1 · p. 3
  • Atria Dawn leads CyberGym with 86.5, 2.0 points above runner-up GLM 5.3 (84.5), scores 86.2 on MLE-bench Lite, and sustains frontier-level competitiveness on ML and software engineering.On GDPval it scores 1583, ahead of DeepSeek V4 Pro (1517), and on JobBench 50.3, ahead of GPT 5.6 sol (45.4).Source: §2.2 · p. 4

2.3. Selected Capabilities and Case Studies

  • Selected discovery cases demonstrate scientific computing, iterative optimization, and sourced literature analysis across distinct research workflows.The weather case processes more than 100 GB and trains a model for 45,000 steps across 69 meteorological variables. GDN optimization reports a 1.46× development latency ratio, not a final benchmark score; a separate report distinguishes a staining readout from causal explanation and proposes follow-up experiments.Source: §2.3 · p. 5
  • Creation and delivery examples range from a persistent MiniOS system and CAD assemblies to sourced reports and presentations organized around task objectives.The MiniOS run takes approximately 20 minutes and checks persistence and autorun across two QEMU sessions; report examples include clean-energy siting, healthcare investment, and semiconductor capital allocation.Source: §2.3 · p. 5
  • In an authorized, isolated environment, Atria Dawn diagnoses and repairs a vulnerability, then re-tests affected paths to externally check the repair.Source: §2.3 · pp. 5–6

3. The Evolution of Human–AI Collaboration

The section describes agents moving from task execution toward planning and iterative project work while humans shift to judgment and intervention; it argues these capabilities alone do not establish self-sustaining RSI.

3. The Evolution of Human–AI Collaboration

  • Agents increasingly plan workflows and revise experimental plans, while human researchers shift from detailed instructions toward higher-level judgment and targeted intervention.The figure depicts this shift from human control, through task running, to partnership around human-defined goals.Source: §3 · pp. 6–7, Figure 5 · p. 7

3.1. From Research Objects to Partners

  • The paper frames three stages: models as research objects, task-level runners within human-led workflows, and project-level coworkers planning and revising scoped R&D under human objectives.Humans retain control of overall workflows and evaluation in the task-runner stage; in the current stage, agents plan and revise within largely human-specified objectives and criteria.Source: §3.1 · p. 7, Figure 5 · p. 7

3.2. Autonomy Still Needs Human Judgment

  • Project-level autonomy is not self-sustaining RSI: agents may execute experiments without identifying valuable next steps or improving their ability to develop successors.Self-sustaining improvement requires agents to identify worthwhile improvements and for successive improvements to strengthen their ability to discover and realize the next.Source: §3.2 · pp. 7–8
  • Researchers prune unpromising searches, counter agents’ correlated blind spots with diverse perspectives, and carry lessons from trial and error across projects to guide what comes next.Their contribution is not performing every experiment, but judging which experiments are worth conducting, which alternatives to keep open, and how experience should shape subsequent exploration.Source: §3.2 · p. 8

4. Human–AI Collaboration in Atria Dawn R&D

AI made a substantial share of reviewed work feasible while humans continued to set goals, choose among options, and steer agent revisions. The records portray collaboration as humans supplying judgment and agents carrying out and revising work.

4. Human–AI Collaboration in Atria Dawn R&D

  • AI was used in 713 of 739 tasks with a clear response (96.5%); the daily median agent-actions-to-human-prompts ratio rose from 11.0 to 28.5 as participant spread widened.Delegated work concentrated on execution and reporting, including train-and-evaluate cycles, debugging, data preparation, resource scaling, and summarizing agent changes.Source: §4 · p. 8, Figure 6 · p. 9

4.1. AI Expands the Range of Feasible Tasks

  • 151 of 455 completed AI-assisted tasks (33.2%) were rated infeasible without AI, across 27 participants, with scope and quality requirements held fixed.These tasks represent work participants said they could not have completed without AI, not simply work that would have taken longer.Source: §4.1 · pp. 9–10, Figure 7 · p. 9

4.2. Role Distribution in Method Proposal and Selection

  • Humans made the final choice in 85.5% of method or parameter decisions and 93.4% of goals or scope decisions; even in AI-infeasible tasks, humans selected the final goal in 95.4% of cases.AI proposed 55.4% of method or parameter options, while its share of final decisions across decision types ranged from 6.1% to 9.2%.Source: §4.2 · p. 10, Figure 8 · p. 11

4.3. Feedback and Revision Across Humans and Agents

  • Among 588 tasks with a recorded difficulty and response, 76.0% moved forward through human intervention; help chiefly involved context or requirement clarification (35.2%) or issue diagnosis or method changes (34.7%).Agents recovered independently in 23.0% of tasks and 1.0% remained unresolved; partial edits and takeover accounted for 3.2% and 0.7% of tasks, respectively.Source: §4.3 · p. 10
  • 95.9% of AI outputs entered deliverables in some form, and agents made revisions after human feedback in 75.4% of substantively revised tasks.Of 627 tasks, 56.5% involved substantive output revisions; humans directly edited in 19.2% of the 354 substantively revised tasks. Revisions often kept the AI draft in place for the agent to rewrite.Source: §4.3 · pp. 10–11, Figure 10 · p. 12

5. Open Challenges

AI’s growing role in R&D does not yet establish sustained recursive self-improvement. The paper identifies technical challenges in research discovery and learning, alongside challenges in preserving meaningful human oversight and safety.

5.1. What technical challenges remain on the path from AI participation in R&D to achieving sustained RSI?

  • Agents proposed options in 55.4% of method and parameter decisions but only 16.9% of goal or scope decisions, and most proposals stayed within researcher-set directions.Greater research volume does not necessarily yield more valuable proposals (Jiang et al., 2025); researchers still prompted agents toward new pipelines and workflows, and models struggle to prioritize directions and promising preliminary findings.Source: §5.1 · pp. 12–13
  • Communication between agents and systematic low-cost exploration could broaden research directions, while anticipation is not necessarily a permanent human advantage.The same agent explored more broadly when researchers supplied their questions, preferences, and trade-offs; anticipation-based pruning may enable decision strategies unlike human intuition (Luo et al., 2025).Source: §5.1 · p. 13
  • Agents often needed researchers to diagnose failures again in later tasks, because lessons remained in sessions or records rather than changing underlying research capability.The paper suggests absorbing experience, potentially through test-time training (Feng et al., 2026a) or architectural redesign, and assessing progress with research-capability-specific tasks (Chan et al., 2025; Zhang et al., 2026). RSI would involve generating more valuable questions and detecting subtler errors across successive rounds.Source: §5.1 · p. 13

5.2. As AI gradually takes over its own R&D, what role can humans play?

  • Humans made the final choice in 85.5% of method and parameter decisions and 93.4% of goal or scope decisions, even as daily median agent actions per prompt rose from 11.0 to 28.5.Researchers retain authority over longer chains of work that one person may not be able to inspect closely; oversight fails if humans cannot assess AI outcomes (Burns et al., 2024).Source: §5.2 · pp. 13–14
  • Researchers must retain responsibility for defining goals and authority to adjust AI R&D, supported by frameworks that help them expand understanding, oversee work, and identify errors and blind spots.The paper calls for clear protocols for authority allocation and alignment standards as AI systems begin to lead most R&D.Source: §5.2 · p. 14
  • Researchers often granted agents broad execution permissions for convenience, but AI should not be given authority to safeguard human safety in the same way as other permissions.The paper calls for proportional investment in alignment and oversight (Greenblatt et al., 2024), with clear guidelines on whose intentions shape alignment, who accepts risks, and who is accountable.Source: §5.2 · p. 14

6. Conclusion

Atria Dawn Preview combines an execution-grounded, externally verified training pipeline with competitive performance across 16 benchmarks. Its development records portray agents as approach proposers and practical executors, while humans primarily contribute judgment and retain final choices.

6. Conclusion

  • Atria Dawn Preview is a 744-billion-parameter MoE agentic language model trained through a Verifiable Experience Pipeline that grounds tasks in real execution and checks outcomes against external signals.The pipeline connects tool-mediated interactions to executable environments and externally verified outcomes.Source: §6 · p. 14
  • Across 16 benchmarks spanning tool use, research, professional work, engineering, and cybersecurity, Atria Dawn shows competitive performance.Source: §6 · p. 14
  • Among 567 reported methods and decisions, AI proposed 64.6%, while humans made the final choice in 85.5%.The development study analyzed 769 task records and agent logs from 56 participants.Source: §6 · p. 14
  • AI was used in 96.5% of 739 tasks with definitive responses regarding AI involvement.Source: §6 · p. 14
  • Human intervention advanced 76.0% of 588 tasks with recorded difficulties, typically through context, clarification, diagnosis, or method adjustments.The findings characterize human contributions primarily as judgment and agents as proposers and practical executors.Source: §6 · p. 14

A.1. Project Lead · B. Evaluation Protocols

The appendix specifies benchmark versions, execution budgets, tools, and scoring procedures for the 16 benchmarks summarized in Table 1.

B. Evaluation Protocols

  • The protocols cover 16 benchmarks and specify their versions, execution budgets, tools, and scoring procedures.The appendix presents these protocols for the benchmarks in Table 1.Source: B. Evaluation Protocols · p. 21
  • AutomationBench uses v1.0.6 with a null-type fix; MLE-bench Lite reports HumanRank scores, and CyberGym uses Claude Code.The AutomationBench version incorporates the fix from PR #13. MLE-bench Lite has a 12-hour evaluation limit and follows the calculation described in Frontis-MA1 (Yang et al., 2026b); CyberGym tasks receive 2 CPUs and 4 GB, with a 4-hour-10-minute timeout and high thinking effort.Source: B. Evaluation Protocols · pp. 21, 23
  • BFCL v4 uses the official harness on 5,217 cases, while its overall score follows weighted category scores and treats API-failure placeholders as incorrect.The harness uses commit 6ea5797, 16 workers, default temperature 0.001, and a 120-second request timeout with at most two attempts; web-search cases use Serper with requests, responses, and search traces recorded. Category weights are 10% Non-Live, 10% Live, 10% Irrelevance, 30% Multi-Turn, and 40% Agentic, with equal weight for Web Search and Memory within Agentic.Source: B. Evaluation Protocols · pp. 21–22
  • SkillsBench averages three runs on a 79-task non-multimodal subset, while τ3-Bench Banking evaluates 97 tasks with a specified retrieval setup.SkillsBench runs through OpenHands in AgentCompass. Banking uses a Tau2/Tau3 harness with BM25, dense retrieval returning the top 10 documents, and a read-only knowledge-base shell; tasks allow up to 200 interaction steps and have a 2,400-second outer timeout.Source: B. Evaluation Protocols · p. 22
  • DeepSearchQA, BrowseComp, and WideSearch use an in-house search, visit, and Python harness; DeepResearch Bench II uses Claude Code and GPT-5.5 scoring.The first three allow up to 500 interaction steps, a 256K-token context window, and 64K output tokens per generation; BrowseComp uses discard-all context management. DeepResearch Bench II gives each task four hours and averages three independent scoring passes.Source: B. Evaluation Protocols · p. 22
  • Workspace-Bench and GDPval use AgentCompass, with task-specific execution limits and LLM-based judging.Workspace evaluations allocate 2 CPUs and 8 GiB per task, use a two-hour limit, and are judged by Doubao Seed 2.0 Lite; GDPval uses OpenHands, 4 CPUs and 8 GB, a four-hour agent timeout, and glm-5.3 pairwise judgments over 220 questions to calculate Elo ratings.Source: B. Evaluation Protocols · p. 22
  • JobBench uses its official OpenCode main split, scores the mean across 65 tasks, and assigns zero to API failures or two failed attempts.Tasks have a 7,200-second timeout, at most two attempts, and are scored by Grok-4.3 after execution.Source: B. Evaluation Protocols · p. 22
  • SWE-bench Pro disables network access and removes Git history, while Terminal-Bench 2.1 reports average pass rate across four runs.SWE-bench Pro uses OpenHands v1.23.0, a 4.5-hour limit, and at most 250 interaction steps. Terminal-Bench uses Claude Code v2.1.224 by default, with a 12× timeout multiplier capped at 12 hours and at most 500 turns; GPT models use Terminus 2, some tasks are refused under built-in safety policies, and the reported Claude Opus 5 configuration falls back to Opus 4.8 for some tasks.Source: B. Evaluation Protocols · p. 23

Related papers

Loading 2609.15818v1…