FARS: An AI System Wrote 100 Research Papers in 228 Hours Without Human Intervention

Forum topic · ✨步子哥 · 2026-02-25

Summary

In February 2025, AI startup Analemma ran FARS (Fully Automated Research System), a live public experiment in which a multi-agent AI pipeline autonomously conducted scientific research for 228 hours 28 minutes with zero real-time human intervention. Built on four cooperating agents—Ideation, Planning, Experiment, and Writing—coordinated through a shared file system and optimized with RLVR (reinforcement learning with verifiable rewards), FARS generated 244 research hypotheses and produced 100 short papers across LLM safety, visual question answering, code generation, and vector database security, processing about 11.4 billion tokens at a total cost of roughly $104,000 (about $1,040 per paper). When evaluated by Stanford's Agentic Reviewer under ICLR standards, the papers averaged 5.05 out of 10—above the 4.21 average of human ICLR 2026 submissions but below the 5.39 average of accepted papers. All code, hypotheses, and papers were published openly on GitLab and the official site. The article analyzes FARS's architecture, compares it with The AI Scientist and other predecessors, and discusses limitations such as incremental-only innovation, peer-review overload, and data-cycle contamination, arguing AI will accelerate rather than replace human research.

FARS: The Lab That Never Sleeps — One AI, 228 Hours, 100 Research Papers

*On the dawn of AI-driven automated scientific research*

Overview

On February 12, 2025, Analemma AI—a startup founded less than a year earlier by Sun Tianxiang, a former core developer of the MOSS large language model, and backed by tens of millions of dollars in angel funding from Sequoia China and Gaorong Capital—launched the first publicly livestreamed fully automated research experiment. Its system, FARS (Fully Automated Research System), ran on 160 NVIDIA GPUs for 228 hours 28 minutes 33 seconds (~9.5 days) with no real-time human involvement, producing:

244 research hypotheses proposed and filtered
100 short papers generated and organized
~11.4 billion tokens processed and generated
Total cost of roughly $104,000 (~$1,040 per paper)

All code, hypotheses, papers, and AI review results were published in real time on GitLab (gitlab.com/fars-a) and the official site (https://analemma.ai/fars).

Key Points

Not a writing tool — an unmanned research factory

FARS is not an AI-assisted writing product. It closes the full research loop—from idea generation to finished paper—using a four-agent multi-agent system:

Ideation: continuously scans open paper databases and GitLab repositories to propose *verifiable* (falsifiable, Popper-style) research hypotheses—244 in total, spanning LLM safety defenses, visual question answering, code generation, and vector database security.
Planning: translates hypotheses into concrete experimental designs (datasets, baselines, comparison groups, metrics).
Experiment: writes its own code, runs experiments on the 160-GPU cluster, debugs failures, and optimizes its strategies using RLVR (Reinforcement Learning with Verifiable Rewards)—using objective outcomes (code runs, metric improvements) as reward signals instead of subjective human ratings, in the spirit behind DeepSeek-R1 and OpenAI's o1/o3.
Writing: assembles hypothesis, methods, results, and analysis into short papers. Short-paper format is deliberate: focus on a single contribution, and encourage reporting negative results as knowledge.

The four agents coordinate through a shared file system—chosen over a vector database for direct handling of structured data (code, logs, drafts), transparency, and easier debugging.

The numbers

Speed: ~2 hours 17 minutes per paper; annualized, theoretically 4,000+ papers versus ~40 from a productive 10-person human team.
Cost: ~$1,040 per paper, versus marginal costs often estimated at thousands of dollars (or tens of thousands of RMB) per paper for human researchers.
Tokens: 11.4 billion in 228 hours—roughly 13,800 tokens per second flowing through reading, coding, analysis, and writing.

How good are the papers?

Analemma evaluated all 100 papers with Stanford ML group's Agentic Reviewer (paperreview.ai) using ICLR's 10-point standard:

| Metric | Score | |:---|:---| | FARS papers, range | 3.0 – 6.3 | | FARS papers, mean | 5.05 | | Mode | ~5.2 (~57 papers) | | ICLR 2026 human submissions, mean | 4.21 | | ICLR 2026 accepted papers, mean | 5.39 |

FARS papers score above the average human submission, slightly below the acceptance line. About 57 papers cluster near 5.2—an "industrial steady state": stable and batch-produced, but lacking breakthroughs. Topics include chart QA under visual degradation (*Re-Inked OCR Views for Robust Chart QA under Visual Degradations*), LLM attack/defense, vector database leakage, quantization, and code generation—solid, incremental contributions with reproducible, fully open code.

A major caveat: AI reviewing AI carries systemic bias risk. The reviewer was trained on human academic norms and may reward formal polish over radical originality. Analemma acknowledges this: at least three researchers with 5+ years of experience manually audit every paper, independent human evaluations are promised, and合格 papers are uploaded to arXiv clearly labeled "AI-generated" for community review.

Not the first, but the furthest

The AI Scientist (Sakana AI, 2024) proved end-to-end feasibility as a proof of concept (arXiv:2408.06292).
ResearchAgent (arXiv:2404.07738) and Agent Laboratory (arXiv:2501.04227) explored LLM-based hypothesis generation and research assistance.
FARS differs by scale and philosophy: 9.5 days of continuous production, fully public, 100 papers, zero real-time human intervention—and an explicit refusal to optimize for top-venue formatting, prioritizing truth-seeking over publication metrics.

Deep concerns

1. Innovation ceiling: Ideation generates hypotheses as a function of existing literature, so output is inherently incremental. Real breakthroughs often require counter-intuitive questioning that lies outside the corpus. 2. Peer-review strain: if such systems proliferate, AI-paper submissions could explode from tens of thousands to hundreds of thousands yearly, overwhelming existing review mechanisms. 3. Data-cycle contamination: AI papers entering training corpora risk recursive self-referential drift, where AI knowledge increasingly derives from AI output rather than human originality.

What remains human

AI cannot (yet) replace: asking genuinely paradigm-shifting questions rooted in embodied cognition; cross-domain insight and paradigm transfer; real-world experiments bridging digital reasoning and physical reality; and ethical judgment about what research should exist. The likely future is a collaboration model where humans act as "PIs of AI PhD teams"—setting vision and value judgments while AI handles heavy literature scanning, trial-and-error, and incremental execution.

Conclusion

As Analemma writes, "This is just the beginning." FARS may mark a phase transition in knowledge production—not the end of research, but its acceleration, in which human effort shifts toward the questions only humans can ask.

References

1. Lu, C., et al. (2024). *The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.* Sakana AI Technical Report. arXiv:2408.06292. 2. Kim, J., et al. (2024). *ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models.* arXiv:2404.07738. 3. Schmidgall, S., et al. (2025). *Agent Laboratory: Using LLM Agents as Research Assistants.* arXiv:2501.04227. 4. Analemma AI. (2025). *FARS: Fully Automated Research System — Official Blog and Live Research Runs Documentation.* https://analemma.ai/fars 5. Ouyang, L., et al. (2022). *Training Language Models to Follow Instructions with Human Feedback.* NeurIPS, 35.

*All FARS run data comes from official public disclosures as of February 23, 2025; review scores come from the official Agentic Reviewer reports. Primary sources: https://analemma.ai/fars and gitlab.com/fars-a.*