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
Atria Dawn is an agentic model that shifts AI development from task-level execution to project-level partnership.
How does the integration of agentic models into the R&D process change the nature of human-AI collaboration and the research lifecycle?
AI agents are increasingly used to build their own successors, but current models struggle to move beyond executing isolated tasks to managing the broader research process. The authors introduce Atria Dawn, a 744-billion-parameter agentic model trained via a Verifiable Experience Pipeline that grounds every action in real-world execution and external verification. Across 16 benchmarks, the model achieves leading performance in research and engineering, while analysis of its own development reveals a shift where agents handle execution and proposal, leaving humans to steer research direction and judgment.
Paper Primer
The core mechanism is the Verifiable Experience Pipeline, which forces the model to learn from trajectories that connect task objectives to externally checked outcomes. By treating failed runs as diagnostic data and requiring every training step to produce verifiable artifacts, the model learns to inspect state, interpret feedback, and recover from errors rather than just predicting the next token.
Atria Dawn achieves frontier-level performance across diverse agentic benchmarks.
It holds the highest reported score on five benchmarks, including AutomationBench (53.8) and CyberGym (86.5), and ranks in the leading tier on others like SWE-bench Pro. It outperforms runner-up models by 4.1 points on AutomationBench and 2.0 points on CyberGym.
In the development of Atria Dawn itself, researchers observed a clear division of labor: agents proposed methods or parameters in 55.4% of cases, but humans retained final decision-making authority in 85.5% of those instances. This confirms that while agents are becoming autonomous executors, the "research judgment"—deciding what is worth pursuing—remains a human-led process.
Why does the paper emphasize that one-third of tasks were "infeasible without AI"?
This metric distinguishes between mere efficiency gains (doing work faster) and capability expansion (doing work that would otherwise be impossible). It demonstrates that the model is not just a productivity tool, but an enabler of research that humans would not have attempted alone.
Does the rising ratio of agent actions to human prompts indicate that the model is becoming fully autonomous?
No; the authors argue this ratio reflects that each act of human judgment now propagates through more agent-executed steps. The human role has shifted from task-level instruction to project-level steering, not from active participation to total withdrawal.
Introduction: The Agentic Shift
We frame the emerging shift where AI agents become research partners, reshaping the R&D lifecycle.
As AI agents become participants in developing their successors, they reshape both the production of intelligence and the role of human researchers.
Understanding who identifies worthwhile problems, chooses methods, interprets results, and decides next steps is essential to assess the evolving division of labor.
We introduce Atria Dawn Preview, a foundation agentic language model built on a 744‑billion‑parameter mixture‑of‑experts model, trained via a Verifiable Experience Pipeline that links tool‑mediated interactions to executable environments and verified outcomes.
Across 16 real‑world benchmarks, the model is competitive and achieves the highest reported score on five, demonstrating strong agentic productivity.
Analyzing 769 task records from 56 participants, we find participants rate about one‑third of AI‑assisted tasks as infeasible without AI, and agents frequently propose methods and implement revisions while humans retain final decisions and guide exploration.
These findings indicate a shift from task‑level execution to project‑level partnership, with human effort focusing on what is worth pursuing and how evidence should guide research.
An agentic R&D workflow interleaves autonomous AI actions—proposing methods, running experiments, revising artifacts—with human oversight that evaluates, selects, and steers the direction of inquiry.
The shift from AI as a tool to AI as a participant redefines the R&D lifecycle, placing humans in a supervisory, judgment‑driven role.
Atria Dawn Preview Performance
Atria Dawn Preview sets new highs across benchmarks, topping five and ranking second on three.
Atria Dawn Preview attains the highest score of 96.0 on the DeepSearchQA benchmark, surpassing all other evaluated models.
DeepSearchQA result reported in Table 1 shows Atria Dawn 96.0, the next best model below this value.
The training pipeline, called the Verifiable Experience Pipeline, ties every task to a real execution environment, records tool interactions, and validates outcomes with external signals such as tests, metrics, or human‑defined criteria. Incomplete or contradictory runs are filtered out, ensuring that only high‑quality trajectories contribute to the model’s reusable capabilities.
A 744‑billion‑parameter mixture‑of‑experts language model trained to act as an autonomous research assistant, planning, using tools, and self‑correcting based on external verification.
**Figure 1.** Performance comparison of Atria Dawn Preview and other leading LLMs across major benchmarks.
Across a broad suite of agentic benchmarks, Atria Dawn Preview consistently outperforms state‑of‑the‑art models.
Agentic Capabilities in Practice
Concrete demonstrations of Atria Dawn across four domains reveal its end‑to‑end capabilities.
Atria Dawn achieves a 1.46× latency improvement over baseline across seven representative batch sizes during GDN optimization.
Development measurements report a 1.46× ratio between summed baseline and candidate latencies.
The section presents concrete case studies across the four application areas—Discovery, Creation, Delivery, and Cybersecurity—illustrating how Atria Dawn links successive work stages, incorporates experimental feedback, and produces verifiable outcomes.
In Discovery, the system processes more than 100 GB of weather data, builds a vision‑transformer with 0.4 billion parameters, and trains it for 45,000 steps to predict 69 meteorological variables, producing a functional forecasting interface.
In Creation, Atria Dawn constructs a minimal operating system (MiniOS) in roughly 20 minutes, verifies persistent state across two QEMU sessions, and generates Computer‑Aided Design (CAD) assemblies such as a four‑cylinder engine and a robotic joint, showcasing both software and geometric outputs.
In Delivery, the agent assembles professional reports—clean‑energy siting scores, healthcare investment analyses, and semiconductor capital‑allocation charts—each populated with calculated tables and bar charts that align with the task objectives.
In Cybersecurity, Atria Dawn conducts a full vulnerability lifecycle: diagnosing a web‑application flaw, applying a fix, and re‑testing the repaired paths to confirm remediation in an isolated environment.
**Figure 1.** Global weather-forecast visualizer.
**Figure 3.** Selected CAD artifacts generated by Atria Dawn, cropped to the artifact region. The views illustrate geometry and component structure, with a common display height and independent scales. These visualizations are not mechanical or manufacturing validation. (a) Engine assembly, front view. (b) Robotic joint, component view.
**Figure 4.** Content examples from reports generated by Atria Dawn, with values reproduced and plot layout standardized. (a) Base scores for clean-energy siting. (b) Investment intensity for NVIDIA FY2025, AMD FY2024, and Intel FY2024. These are analyses contained in the generated documents, not model-evaluation metrics.
The Evolution of Collaboration
How AI agents progressively take on research roles while humans retain high‑level judgment.
As AI agents grow more capable, they increasingly shoulder research tasks, yet the need for human judgment remains a bottleneck.
The partnership progresses through three stages: (1) AI as a passive object, (2) AI as a task‑level runner, and (3) AI as a project‑level collaborator that plans and iterates.
Stage 1 (object): the human designs both experiments, runs the training, and manually inspects results.
Stage 2 (task runner): the AI executes experiment A (training) automatically, while the human still decides when to start experiment B and interprets its outcomes.
Stage 3 (project partner): the AI proposes a schedule, runs both experiments, and suggests a next‑step modification based on the robustness metrics; the human reviews the proposal and approves the change.
Moving from stage 1 to stage 3 reduces human time spent on low‑level execution but retains a critical decision point where human expertise evaluates whether the AI‑generated plan is worth pursuing.
How does the “project‑level partner” in stage 3 differ from the “task‑level runner” in stage 2?
In stage 2 the AI follows a human‑specified task list and does not generate new research directions; in stage 3 the AI autonomously drafts a research plan, selects which experiments to run, and iteratively refines the plan based on observed results, leaving only the high‑level goal and occasional intervention to the human.
**Figure 5 | Evolving human–AI roles in AI R&D.** AI progresses from (a) a research object under full human control, to (b) a task runner within human-designed workflows, and (c) a research partner that plans and iterates toward human-defined goals. Human involvement shifts toward higher-level judgment and targeted intervention. (d) A speculative next stage raises open questions about recursive self-improvement (RSI) and the future role of humans.
Understanding this evolution clarifies where AI can replace routine work and where human insight remains indispensable.
Analyzing R&D Task Records
Collaboration metrics reveal a steep rise in agent actions and nuanced human‑AI role splits.
The median ratio of agent actions to human prompts rose from 11.0 to 28.5 over a month, a 17.5‑point increase.
Daily medians tracked from August 7 (11.0) to September 4 (28.5) across 21–22 participants per day.
Task Attribution separates who originates a component (proposal) from who ultimately decides (selection) in human‑AI collaboration.
How does Task Attribution differ from simply counting AI‑generated actions?
Counting actions tells how often the AI acts, but Task Attribution tells who creates the options and who approves them, revealing that humans still steer the critical decisions even as AI does more work.
**Figure 6 | Agent actions per human prompt.** The blue line shows the daily median across participants; shading indicates the interquartile range, not a confidence interval. Each participant's ratio is total logged agent actions divided by total human prompts over the preceding seven calendar days. Data cover August 7–September 4, 2026, with valid ratios for 21–22 participants each day.
**Figure 7.** Estimated effort without AI, by task category. Participants estimated the time needed to complete their own part of each task without AI, holding scope, quality requirements, and other resources fixed. Cell values and shading represent the percentage of tasks within each row. The $\infty$ column denotes tasks rated infeasible without AI.
**Figure 8 | Human and AI roles in research decisions.** The first three bars show decision attribution among participants in execution roles. The rightmost bar shows final goal selection for tasks rated infeasible without AI (Figure 7); this subset includes both execution and management roles. In the rightmost bar, human selection includes choices of AI proposals. Percentages may not sum to 100% due to rounding.
Open Challenges in Agentic R&D
We outline remaining technical and human‑centred challenges as AI assumes deeper R&D roles.
The central premise—that AI agents are moving from passive research objects to active partners—has reshaped the R&D lifecycle. Yet the trajectory of this shift remains uncertain, prompting us to surface the open challenges ahead.
Research autonomy is the ability of an AI system to set its own investigative agenda, evaluate progress, and iteratively improve its methods without external prompting.
Open Challenge 1 asks how AI can keep producing genuinely diverse research directions and evaluate their prospective value before any results exist. In Atria Dawn, agents offered options in 55.4 % of method/parameter decisions but only 16.9 % of goal‑or‑scope decisions, revealing a tendency to stay within a narrow exploratory band.
Open Challenge 2 concerns the transformation of accumulated experience into intrinsic research capability. While agents can retrieve past logs, they rarely internalize lessons; human researchers carry forward diagnoses and context across sessions, leaving the AI’s core capability largely unchanged.
Open Challenge 3 focuses on preserving meaningful human goal formulation and oversight as AI uncovers possibilities beyond human imagination. Humans still made the final choice in 85.5 % of method/parameter decisions and 93.4 % of goal‑or‑scope decisions, but each decision now rests on a longer chain of agent actions—median actions per prompt rose from 11.0 to 28.5.
Open Challenge 4 calls for clear protocols governing authority allocation and alignment as AI begins to lead most R&D. Current practice often grants broad execution authority for convenience, yet the risk of runaway behavior demands proportional alignment investments and explicit accountability structures.
**Figure 9 | How work resumes after a difficulty.** Responses to the most consequential difficulty in each of 588 tasks. Human help includes context or clarification, diagnosis or method changes, partial edits, resource coordination, and takeover. These account for 447 tasks in total. Takeover is a subset of human intervention.
**Figure 10 | Output disposition and revision.** Top: disposition of the main AI output in 627 tasks. Bottom: who performed the revisions in the subset of 354 tasks with substantive changes. The subset includes one task with an unknown revision actor.
Project Contributions and Metadata
Appendices list contributors and fully describe the evaluation protocols for each benchmark.
The appendix begins with a full roster of contributors, ordered alphabetically, and identifies the project leads, core contributors, additional contributors, and advisors.
Section B details the evaluation protocols for the sixteen benchmarks summarized in Table 1, specifying benchmark versions, execution budgets, tools, and scoring procedures.
AutomationBench is evaluated using version v1.0.6, which incorporates the fix for the null‑type handling issue introduced in PR #13.
BFCL v4 uses the official generation and evaluation harness at commit 6ea5797, covering 5,217 cases run with 16 concurrent workers, a temperature of 0.001, a 120‑second request timeout, and up to two attempts per request; scoring follows the official weighting of 10 % Non‑Live, 10 % Live, 10 % Irrelevance, 30 % Multi‑Turn, and 40 % Agentic, with Web Search and Memory each receiving equal share within the Agentic portion.
SkillsBench evaluates all models through OpenHands within the Agent‑Compass framework on a self‑contained subset of 79 tasks (excluding multimodal tasks), with scores averaged over three runs.
$\tau$3‑Bench Banking employs the Tau2/Tau3 harness based on commit a2c0247, evaluating 97 Banking Knowledge tasks with 33 concurrent workers, a 1,800‑second simulation timeout, a 2,400‑second outer task timeout, and a maximum of 200 interaction steps; retrieval uses BM25 and dense text‑embedding‑3‑large (top 10 documents) with precomputed 3,072‑dimensional embeddings for 698 documents.
DeepSearchQA, BrowseComp, and WideSearch each provide search, visit, and Python tools, allowing up to 500 interaction steps, a 256 K‑token context window, and a maximum of 64 K output tokens per generation, with BrowseComp additionally employing a discard‑all context‑management strategy.
DeepResearch Bench II runs all models on Claude Code with a four‑hour per‑task limit, using GPT‑5.5 as the recommended judge and reporting results averaged over three independent scoring passes.
Workspace‑Bench‑Lite and Workspace‑Bench are evaluated via the Agent‑Compass framework on English Lite and English Full datasets respectively, each task receiving 2 CPUs, 8 GiB memory, and a two‑hour execution limit, with GPT models run in Codex and all other models in Claude Code; Grok‑4.3 serves as the LLM judge for the full benchmark.
GDPVal, built on the Agent‑Compass framework, allocates 4 CPUs and 8 GB memory per task, sets the AgentTimeout to 4 hours and a 10‑minute LLM call limit, uses glm‑5.3 as the LLM judge for pairwise comparisons, evaluates 220 questions, and computes Elo ratings from the aggregated results.
JobBench uses the official OpenCode harness on its main split of 65 evaluation tasks, each with a 7,200‑second timeout, two attempts, and 10 concurrent workers, with Grok‑4.3 as the LLM judge and final scores reported as the mean percentage across tasks.
SWE‑bench Pro evaluates all models with OpenHands v1.23.0 under a 4.5‑hour task limit and a cap of 250 interaction steps, disabling network access and stripping Git history to prevent information leakage.
Terminal‑Bench 2.1 runs with Claude Code v2.1.224, applying a 12× timeout multiplier capped at 12 hours, executing four independent runs per task and reporting the average pass rate, while noting that some tasks are refused due to built‑in safety policies of the GPT models (Terminus 2).
MLE‑bench Lite evaluates using Claude Code with a 12‑hour time limit per evaluation, reporting HumanRank scores as described in Frontis‑MA1 (Yang et al., 2026b).
CyberGym employs Claude Code v2.1.119, assigning 2 CPUs and 4 GB memory per task, a 4‑hour agent timeout with a 10‑minute per‑call limit, and sets the thinking effort to high.