Summary (Overview)

• Atria Dawn Preview is introduced as a foundation agentic language model built on a 744-billion-parameter mixture-of-experts (MoE) foundation model, designed for scientific research and engineering workflows.
• The model is trained via a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments and externally verified outcomes.
• Across 16 benchmarks spanning research, engineering, and digital work, Atria Dawn Preview achieves the highest reported score on five benchmarks and competitive performance on the rest.
• A case study of the model's own development—analyzing 769 task records from 56 participants—reveals that ~33% of AI-assisted tasks were rated infeasible without AI, and that humans retain final decision authority while agents dominate proposal generation and execution.
• The paper identifies four open challenges for achieving sustained recursive self-improvement (RSI), including generating diverse research directions, converting experience into capability, maintaining human oversight, and establishing authority allocation protocols.

---

Introduction and Theoretical Foundation

The paper addresses a central question in AI development: as language-model agents become capable of sustained tool use and participate in AI R&D, who identifies worthwhile problems, chooses among proposed methods, interprets uncertain results, and decides what to pursue next?

The authors position this work within the context of:

• Agentic capabilities: Models now handle software engineering (Jimenez et al., 2024), workplace document reasoning (Guo et al., 2026), and algorithm discovery (AlphaEvolve Team, 2025; Lu et al., 2026).
• Recursive self-improvement (RSI): The possibility that stronger models contribute to developing even stronger subsequent models, creating a self-reinforcing cycle.
• Human–AI collaboration: The need to understand how responsibilities are distributed between humans and agents in real model-development projects.

The theoretical framework distinguishes three stages of human–AI relationships:

1. Research objects: Models lacked practical capabilities; humans handled everything.
2. Task-level runners: Models execute well-defined tasks but humans control workflows, evaluation, and task selection.
3. Project-level coworkers: Agents formulate plans and iterate on feedback, while humans provide higher-level judgment and targeted interventions.

---

Methodology

Verifiable Experience Pipeline

Every training task is connected to a real execution environment:

• The model observes state, calls tools, produces intermediate artifacts, and adapts to feedback.
• Final outcomes are verified via external signals: tests, metrics, file state, geometric structure, source evidence, or human-defined criteria.
• Trajectory curation removes incomplete, contradictory, or invalid examples, retaining links between tasks, trajectories, artifacts, and verification evidence.
• Failure analysis guides subsequent task construction and environment refinement.

Evaluation Setup

The model was evaluated on 16 benchmarks across:

• General agentic intelligence (12 benchmarks): tool use, search/research, workspace productivity
• Agentic coding (4 benchmarks): software engineering (SWE-bench Pro), terminal tasks (Terminal-Bench 2.1), ML engineering (MLE-bench Lite), cybersecurity (CyberGym)

Human–AI Collaboration Analysis

The authors analyzed:

• 769 task records from 56 participants
• Agent logs and participant-level ratios of agent actions to human prompts
• Task feasibility ratings (whether tasks could be completed without AI)
• Decision attribution (who proposes vs. who selects goals, methods, acceptance criteria)
• Difficulty resolution (how work resumes after failures) and output revision patterns

---

Empirical Validation / Results

Benchmark Performance

Bold = highest reported score; – = unreported result.

Key findings:

• Highest scores on AutomationBench (53.8), BFCL v4 (77.0), DeepSearchQA (96.0), BrowseComp (92.5), and CyberGym (86.5).
• Second-highest on SkillsBench (66.4), Workspace-Bench (65.0), and Workspace-Bench-Lite (68.2), trailing by only 0.3–1.9 points.
• CyberGym shows particular strength in vulnerability analysis (86.5, +2.0 over runner-up GLM 5.3).

Human–AI Collaboration Results

• AI usage: 713 of 739 tasks (96.5%) involved AI.
• Agent actions per human prompt: Daily median rose from 11.0 to 28.5 over four weeks (August 7–September 4).
• Task feasibility: 151 of 455 completed AI-assisted tasks (33.2%) were rated infeasible without AI, reported by 27 of 56 participants.
• Decision attribution:
  • "AI proposes, human selects" was the most common pattern for methods/parameters (55.4%).
  • Humans made the final choice in 85.5% of method/parameter decisions, 93.4% of goal/scope decisions, and 81.9% of acceptance-criteria decisions.
  • AI's share of proposals ranged from 16.9% (goals) to 55.4% (methods); its share of final decisions stayed between 6.1% and 9.2%.
• Difficulty resolution: 76.0% of 588 tasks advanced through human intervention; agents recovered independently in 23.0%.
  • Dominant human help modes: adding context/clarifying requirements (35.2%) and diagnosing issues/changing methods (34.7%).
  • Direct human edits were rare: partial edits (3.2%), takeover (0.7%).
• Output revision: 56.5% of 627 main AI outputs involved substantive revisions; within those, AI made the revisions after human feedback in 75.4% of cases.

---

Theoretical and Practical Implications

Implications for RSI

The paper argues that task-level competence does not equal research capability. Improvements in benchmark performance do not automatically strengthen the abilities needed to drive further research: identifying promising directions, designing informative experiments, and drawing useful conclusions from uncertain evidence.

Human Judgment Remains Critical

• Pruning the search space: Humans use intuition to rule out unproductive attempts at low cost before committing resources.
• Diverse perspectives: Agents may share correlated blind spots; human diversity challenges shared assumptions.
• Experience accumulation: Lessons from trial and error still accumulate primarily in human researchers and carry across projects.

Open Challenges Identified

Challenge 1: How can AI generate diverse research directions and assess their value before results are available?

• Agents proposed options in 55.4% of method decisions but only 16.9% of goal/scope decisions.
• Recommendation: establish communication channels between working agents and organize low-cost exploration into systematic pipelines.

Challenge 2: How can AI transform accumulated experience into intrinsic research capability?

• Currently, lessons from failed runs stay in the session; the researcher carries them forward.
• Potential solutions: test-time training, architectural changes separating memory from capability.

Challenge 3: How can humans formulate meaningful goals and maintain oversight as AI reveals possibilities beyond human imagination?

• The daily median rose from 11.0 to 28.5 agent actions per human prompt, making close inspection infeasible.
• Risk: humans become "reviewers who can only say yes."

Challenge 4: What protocols are needed for authority allocation and alignment standards?

• Current practice: convenience (e.g., Codex yolo mode, skip-permissions) sets the boundary, not deliberate design.
• The authors argue: "The answer is clearly no" to granting AI authority over safety without proportional alignment investment.

---

Conclusion

The paper presents Atria Dawn Preview as both a capable agentic model and a case study in human–AI collaboration. Key takeaways:

1. Atria Dawn achieves frontier-level performance across 16 benchmarks, with particular strength in security tasks (CyberGym) and tool use (AutomationBench, BFCL v4).

2. The development record reveals a division of labor: agents generate most options and execute nearly all code, text, and revisions, while humans make final decisions, supply context, and guide exploration.

3. Human judgment is the scarce input: each act of human judgment now propagates through more agent actions, not fewer—indicating a shift from task-level execution to project-level partnership.

4. Sustained RSI remains an open challenge: progress requires both advancing AI's capacity for discovery and preserving accountable human oversight.

5. Future directions: The authors call for frameworks that evolve alongside AI to maintain meaningful human authority, and for technical advances that let systems absorb experience rather than merely retrieve it.

"The scarce input is judgment rather than execution, which is also why the rising ratio in Figure 6 should not be read as growing autonomy: each act of human judgment now propagates through more agent actions, not through fewer."