*— On FARS and the dawn of the AI-driven research revolution*
---
🌙 Prelude: Deep in the night, something kept writing papers
On the morning of February 12, 2025, in an ordinary server room, 160 NVIDIA GPUs lit up simultaneously. No PhD students staying up late revising manuscripts, no professors red-marking experimental data, no group meetings, no humming coffee machine — only code flowing quietly, like current through the deep sea.
That day, an AI company called Analemma pressed the "start button," launching humanity's first publicly livestreamed fully automated research experiment. Founded less than a year earlier, the company had raised tens of millions of dollars in angel funding from Sequoia China, Gaorong Capital, and others; its founder, Sun Tianxiang, was a core developer of the MOSS large language model. They gave the system an ambitious name: FARS — Fully Automated Research System.
Nine and a half days later, on February 23, the timer stopped at 228 hours, 28 minutes, and 33 seconds. Without any real-time human intervention, the machine had, entirely on its own:
- Proposed and filtered 244 research hypotheses
- Generated and organized 100 short papers
- Processed and generated 11.4 billion tokens
- Cost roughly $104,000 in total
- Visual question answering enhancement (*Re-Inked OCR Views for Robust Chart QA under Visual Degradations*)
- LLM safety defenses and attacks
- Vector database information leakage
- Model quantization optimization
- Code generation quality improvement
- Not a one-off demo, but 9.5 days of continuous production
- Not a closed test, but a fully public livestream
- Not a small sample, but a systematic output of 100 papers
- Not human-assisted, but zero real-time human intervention
This was not science fiction, nor a secret black-box experiment — all code, hypotheses, papers, and AI review results were publicly available in real time on GitLab (gitlab.com/fars-a) and the official website, open for anyone to observe.
But what does this actually mean? Is AI really "doing research"? Are the papers any good? Do human researchers still have a future?
With these questions, let us enter this laboratory that never sleeps.
---
🏗️ Blueprint of a research pipeline: Not a "paper-writing tool"
To understand FARS, we first need to dispel a common misconception: it is not an AI-assisted writing tool.
If you have seen AI products that polish articles or generate summaries, set that impression aside for now. FARS is more like an unattended research factory — it finds its own topics, designs its own experiments, writes and runs its own code, and organizes the results into papers on its own. The entire research loop, from "having an idea" to "writing a paper," is completed independently by the machine.
The core of this factory is a four-agent multi-agent system.
> 📌 Concept note: What is a "multi-agent system"? > > In AI, an "agent" is a program that can perceive its environment, make decisions, and act autonomously. A multi-agent system lets multiple agents collaborate, each specializing in a different subtask, working in relay toward a complex goal — like different departments in a company, each with its own responsibilities yet interdependent.
An intuitive metaphor: imagine an elite research squad with four specialists — an investigator, a planner, an experimenter, and a writer — each with clear duties, jointly completing every project. That is how FARS's four agents work:
🔍 First: Ideation
The Ideation agent is the pipeline's "creative engine." It continuously scans the latest open paper databases and GitLab code repositories, understands frontier research hotspots, and proposes verifiable research hypotheses.
Note the word "verifiable" — it is central to FARS's design philosophy. Ideation does not brainstorm freely; every hypothesis it proposes must be falsifiable or confirmable by experiment. This principle traces back to Karl Popper's classic idea that a meaningful scientific proposition must be falsifiable. Over the 228-hour run, Ideation generated 244 hypotheses spanning LLM safety defenses, visual question answering, code generation optimization, vector database security, and other AI frontier topics.
📐 Second: Planning
The planner translates abstract hypotheses into concrete experimental protocols. It must answer: which datasets? Which models as baselines? Which control groups? What metrics?
This step looks simple but is critical. A good experimental design is like a well-prepared debate — you must anticipate every possible rebuttal in advance and design the means to counter it. The Planning agent takes on "experiment design," a core part of human research that depends heavily on experience.
⚗️ Third: Experiment
This is FARS's most "hardcore" member. The Experiment agent writes its own code, runs experiments on the 160-GPU NVIDIA cluster, collects results, and analyzes data.
It does not merely call pre-written scripts — it writes experimental code from scratch according to the Planning agent's protocol, handles errors and exceptions, and debugs and adjusts when results deviate from expectations. Even more impressively, this agent uses RLVR (Reinforcement Learning with Verifiable Rewards) to optimize its own strategies — a secret weapon we will detail below.
✍️ Fourth: Writing
After the experiment, the Writing agent takes the baton, organizing the whole research process — hypothesis, methods, results, analysis, conclusion — into a structurally complete short paper.
Analemma deliberately chose the "short paper" format rather than the traditional lengthy academic paper. Behind this lies a deep design philosophy: focus on a single contribution, and encourage reporting negative results. Rather than stuffing pages with related work, a few pages should clearly state: what I did, what I found, what I concluded — including "failed experiments," because negative results are also knowledge.
---
🧬 Shared file system: The brain center of the four agents
How do the four agents collaborate? The answer is surprisingly simple: through a shared file system.
In FARS's architecture, all agents read and write to the same file system — it stores literature summaries, experiment logs, code files, intermediate results, paper drafts, and more. It is both the "workbench" for passing information between agents and the system's "long-term memory." Ideation writes a hypothesis, Planning adds an experimental protocol to the same file, Experiment reads the protocol and writes in results, and Writing integrates everything into a paper. This pipeline runs on a shared folder.
> 📌 Concept note: Why a file system instead of a vector database? > > Many AI systems use a "vector database" — converting text into mathematical vectors and retrieving related content via similarity computation, like a super-smart search engine. FARS chose the more direct shared file system. This is an engineering-wise choice: file systems handle structured data (code, logs, drafts) more directly, agents can "share documents" like a real team, intermediate abstraction layers cause less information loss, and the whole process stays more transparent and easier to debug.
The design recalls how real research teams work — professors and PhD students sharing folders on a lab server, with drafts, data, and scripts visible and editable by everyone. FARS's four agents "team up" in this plain yet effective way.
---
⚡ The secret weapon: What is RLVR, and why does it matter?
One keyword appears repeatedly in FARS's tech stack: RLVR — Reinforcement Learning with Verifiable Rewards.
To understand RLVR, start with its predecessor, RLHF (Reinforcement Learning from Human Feedback). Over the past few years, RLHF has been widely used to train conversational models like ChatGPT: human evaluators score AI responses, and those scores train the model to better match human preferences.
RLHF has a fundamental limitation: it depends on human scoring, which is subjective, fatiguing, and costly — you cannot have humans evaluate, in real time, every decision an AI makes while conducting research.
RLVR solves this. Its core idea: use objective, verifiable outcomes as the reward signal, replacing subjective human scores.
> 📌 Concept note: Reinforcement learning > > Reinforcement learning is a branch of machine learning built on rewards and punishments: an agent tries actions in an environment, receives rewards or penalties based on outcomes, and gradually learns good strategies — like training a dog with treats. RLVR's innovation is turning the "reward" from subjective human ratings into objective, verifiable experimental outcomes: code runs and metrics improve — reward; crashes and off-target results — penalty.
In research settings, "verifiable rewards" exist naturally: did the code run? Did the metrics improve? These outcomes are objective, and the machine can judge them itself. RLVR thus allows the Experiment agent to continuously optimize its experimental strategies through self-reinforcing learning — success is reward, failure is a lesson, round after round, self-evolving.
This technique rapidly became an AI research hotspot in 2024–2025. DeepSeek-R1 and OpenAI's o1/o3 series all heavily adopted RLVR-related ideas for reasoning. Bringing RLVR into automated research pipelines, as FARS does, is a valuable engineering translation of frontier academic work.
---
📊 The 228-hour marvel: The staggering numbers
Let us examine the striking numbers and try to grasp what they truly mean.
Speed: a paper every 2 hours
In 228 hours 28 minutes 33 seconds — about 9.5 days — FARS produced 100 papers, an average of about 2 hours 17 minutes per paper.
A human PhD student typically needs months to half a year, from topic selection to manuscript, for one paper. Even fast conference submissions take weeks from project launch to deadline. A 10-person team where each person writes one paper per quarter would be highly productive at 40 papers a year; FARS's annualized output theoretically exceeds 4,000 papers.
Cost: a knowledge factory at $1,040 per paper
FARS's total cost was about $104,000 — roughly $1,040 per paper.
In isolation, that sounds high. By comparison: the marginal cost per first-author paper from a PhD student at a Chinese university (tuition, stipends, advisor time, equipment depreciation) is often estimated in the tens of thousands of RMB; top Silicon Valley lab researchers often earn $200–500 per hour, so a single paper's labor cost easily exceeds $10,000.
The price signal is clear: the marginal cost of AI-driven research is rapidly approaching a threshold that is hard to ignore.
Token flood: a spectacle of 11.4 billion
What does 11.4 billion tokens mean?
GPT-4's training data is estimated at roughly 1 trillion tokens; FARS inferred and generated about 11.4 billion tokens in 228 hours — about 13,800 tokens flowing through the system per second. Printed on A4 paper, these tokens would fill a medium-sized classroom.
What did these tokens flow through? Literature reading, hypothesis generation, experiment code writing, data analysis, paper writing, self-review... every research activity in this digital factory became a surge and transformation of tokens.
---
🔬 AI reviewing AI: How good are those 100 papers?
The quantity is fascinating, but the more crucial question: are the papers any good?
Analemma used the Agentic Reviewer developed by Stanford's ML group (paperreview.ai) to evaluate all 100 papers uniformly, scored on ICLR conference review criteria out of 10.
> 📌 Concept note: What is ICLR? > > ICLR (International Conference on Learning Representations) is one of the top AI academic conferences, alongside NeurIPS and ICML — together the "big three." Annual acceptance rates run about 25–30%; reviews use a 1–10 scale, and scores above 6 are generally considered strong acceptance prospects.
The results:
| Metric | Value | |:---------|:-----| | FARS paper score range | 3.0 – 6.3 | | FARS paper average | 5.05 | | FARS paper modal cluster | ~5.2 (about 57 papers concentrated here) | | ICLR 2026 human submissions overall average | 4.21 | | ICLR 2026 accepted papers average | 5.39 |
These numbers deserve careful reading.
5.05 vs 4.21: FARS's average score is higher than the overall level of human submissions. In other words, a paper drawn at random from FARS's output would, in expectation, be better than a random human submission.
5.05 vs 5.39: Compared with accepted human papers, FARS's average is slightly lower — a gap of about 0.34 points from the acceptance line, not yet consistently meeting top-conference standards. But the gap is small; in the distribution's right tail, some FARS papers scored above 6, touching the acceptance range.
Structural analysis: about 57 papers clustered around 5.2 — an interesting "industrial steady state," with mass output concentrated in one quality band and a relatively small standard deviation. This is exactly what pipeline production looks like: stable, batch-produced, but lacking breakthroughs.
From the publicly available titles, the content covers:
These are all "mainstream" AI research directions — solid but not radical incremental contributions. Each has a clear problem statement, a reasonable method, reproducible experiments — no "reinventing the Transformer" level of disruptive work, but no hollow padding either. Code is fully open-sourced and experiment logs transparent, so the community can independently reproduce and verify.
That said, we must put an important asterisk on these numbers: AI reviewing AI carries a risk of systematic bias. The Agentic Reviewer's scoring is trained on the writing norms of human academic papers — it may score "well-structured, properly formatted" papers higher while under-rating radical, unconventionally expressed original work. A deeper concern: FARS and the Agentic Reviewer have, in some sense, "learned" from the same corpus of human papers — if FARS's Writing agent produces conventionally formatted articles, the reviewer will naturally give high scores, but this may be merely resonance at the format level, not genuine recognition of scientific value.
Analemma does not shy away from this, and additionally arranged for at least 3 researchers with 5+ years of experience to manually review every paper, promising independent human quality assessments soon. Qualified papers, clearly labeled "AI-generated," are uploaded to arXiv — not submitted through traditional conference channels — welcoming peer review from the global research community.
---
🌐 Not the first, but the farthest along
FARS is not the first attempt at "AI doing research automatically," but it has taken the longest step yet.
In 2024, Japan's Sakana AI released a milestone work: The AI Scientist [1]. It could autonomously propose research ideas, write code, run experiments, generate complete papers, and perform automated review — in small-scale tests it produced several papers at NeurIPS Workshop acceptance level, sparking wide academic discussion. But AI Scientist's operating scale was limited, more a proof of concept demonstrating the technical feasibility of end-to-end automated research.
That same year, research teams from Stanford and Carnegie Mellon released ResearchAgent [2] and Agent Laboratory [3], exploring how large language models could iteratively generate research ideas, review literature, and design experiments. These systems showed impressive capabilities on specific tasks but none built a complete end-to-end industrial pipeline.
FARS's innovation lies in achieving industrial-scale continuous operation:
Another dimension of comparison is research philosophy. Many automated research tools aim at "publishing top-conference papers," aligning generated papers' format and style with conference standards. FARS goes the other way, explicitly stating it does not optimize for conference formats, producing short papers and encouraging negative results. This "first-principles" design is closer to the true spirit of science: the purpose of research is to explore truth, not merely to publish papers.
---
🤔 Deep concerns we must face
In the face of FARS's achievements, we cannot simply marvel. Some deep problems must be raised seriously.
The ceiling of innovation: the inherent limits of incremental work
Among the 100 papers, there is almost no truly disruptive work. This is no accident.
FARS's Ideation agent generates hypotheses by reading existing papers. Its creative space is essentially a function of the existing literature space — the best it can do is find gaps in current research and fill them incrementally.
True scientific breakthroughs often come from counter-intuitive insight, cross-domain inspiration, or daring to question out of ignorance. A scholar who has read every 19th-century physics paper would not necessarily have independently invented quantum mechanics — because that breakthrough required questioning the "common sense of classical physics," and such questioning often comes from the naivety of an outsider, not the erudition of a scholar.
Pressure on the academic ecosystem: can peer review cope?
As AI research tools spread, a practical question grows urgent: when AI can rapidly generate large numbers of reproducible but incremental papers, how does academia maintain standards for judging truly original work?
Journal editors, review committees, funding panels — these gatekeeping mechanisms rest on the assumption that "paper output is limited, human review is feasible." If systems like FARS become widespread and AI-paper submissions in a field explode from tens of thousands to hundreds of thousands per year, the existing peer-review system may face fundamental strain. Who will read these papers? Who judges their true value?
Data-loop contamination: the long-term risk of AI training on AI
There is a deeper worry: if AI-generated papers enter academic databases at scale, and the next generation of AI is trained on them, a closed data loop forms. Over time, AI's "knowledge" increasingly derives from AI's own output rather than original human thought. Like a photocopier repeatedly copying copies — each generation accumulates error — this could cause a systematic drift in scientific knowledge, shaped more by AI's expressive habits than by the real world's laws.
This is not unique to FARS, but FARS's scale makes the problem more urgent and more deserving of serious thought by the global scientific community.
---
🚀 The future of research: where do humans go?
So, back to the most fundamental question: when AI can produce a competitive paper every 2 hours, do human researchers still matter?
The answer is yes — and for more reasons than you might expect.
Human irreplaceability: what AI cannot do alone
Asking genuinely disruptive questions. The most important moments in the history of science were not when answers were found, but when the right questions were asked. Why do apples fall? Is the speed of light independent of the observer's motion? Does DNA carry hereditary information? Asking these requires deep judgment about "what is worth asking" — judgment rooted in humans' embodied cognition and lived experience, which AI cannot yet replicate.
Cross-domain insight and paradigm transfer. A neuroscientist attends a concert and suddenly sees an analogy between the brain's rhythmic activity and musical rhythm, spawning a whole new research direction. Such flashes require human sensory experience, aesthetic judgment, and implicit cross-domain connections. AI can find textual cross-domain associations in existing literature but rarely produces genuinely "paradigm-breaking" new understanding.
Real-world experiments and validation. Everything FARS does happens in the digital world — text, code, model weights. But science does not live only on screens. Drug trials need human participants; materials science needs real atomic arrangements; climate models need calibration against the real Earth system. Human researchers remain the indispensable bridge between digital reasoning and physical reality.
Ethical judgment and social meaning. Which research deserves doing? Which technologies should not be developed? AI can produce "efficiency-optimal" research strategies but cannot answer, for humans, whether something benefits humanity. Research's social responsibility, ethical review, and value judgments about "why this matters" are territory humans must hold.
A new human-AI paradigm: you become the boss of an AI PhD team
A more hopeful outlook: systems like FARS will free human researchers from heavy repetitive work, letting them focus on higher-level creative work.
Imagine a near-future research scenario:
> A human scientist holds a grand but vague research vision — "I want to understand how the brain maintains cognitive stability in noisy environments." She summons a team of AI agents and authorizes them to scan relevant literature, propose testable sub-hypotheses, and design baseline experiments. Days later, the AI team returns 20 preliminarily validated research directions, with data and code. The scientist reviews these results and, with her professional intuition and domain insight, selects the 3 most valuable directions for deep human-AI collaborative research.
In this scenario, AI handles "heavy surveying and trial-and-error," humans handle "strategic direction and value judgment" — the optimal division of labor.
This mirrors calculators, computers, and the internet: the stronger the tools, the more human energy goes to genuinely creative, meaningful work. Just as programmers with GitHub Copilot write code faster, making architecture design, product judgment, and user understanding more important, future researchers will need to invest more in big-question thinking, cross-domain integration, and critical evaluation of experiments — and can achieve bigger breakthroughs because of it.
Research will not disappear — it will only accelerate, amplify, and become more interesting.
---
🌅 Epilogue: this is only the prelude
In February 2025, FARS wrote 100 papers in 9.5 days. And then... the system kept running.
Analemma wrote on its official blog: "This is just the beginning." They plan to keep iterating the system, improving each agent's capabilities and paper quality, and inviting the global research community to review and give feedback.
Throughout humanity's long history of knowledge accumulation, every revolution in production tools has triggered deep anxiety — the printing press displaced monk scribes, computers displaced typists, the internet eliminated many traditional jobs. But after every revolution, humanity's knowledge production leaped to new heights, giving rise to previously unimaginable new professions and possibilities.
FARS's emergence may be a similar moment in the history of research. It is not the end of science, but a qualitative leap triggered by quantitative change — when the incremental production of knowledge is massively accelerated, human energy can flow increasingly into what only humans can do: asking big questions, defining future directions, and giving knowledge meaning.
That server room with its never-extinguishing lights is not replacing scientists' laboratories — it is preparing a faster, stronger toolbox for scientists.
As for you, reading this article: if you love research, love exploration, love asking "why" — this era needs you more than ever.
Only now, your postdoc team might be made of AI.
---
📚 References
1. Lu, C., et al. (2024). *The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.* Sakana AI Technical Report. arXiv:2408.06292. > The first complete AI autonomous research system, covering hypothesis generation, experimentation, paper writing, and review end to end, validating the technical feasibility of fully automated research.
2. Kim, J., et al. (2024). *ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models.* arXiv:2404.07738. > An LLM-based system for iterative research idea generation, exploring how multi-round feedback improves hypothesis quality.
3. Schmidgall, S., et al. (2025). *Agent Laboratory: Using LLM Agents as Research Assistants.* arXiv:2501.04227. > A systematic framework for LLM agents as research assistants, discussing multi-agent collaboration in laboratory settings.
4. Analemma AI. (2025). *FARS: Fully Automated Research System — Official Blog and Live Research Runs Documentation.* https://analemma.ai/fars > Official technical documentation and experiment records for FARS, including full metadata for the 100 papers, review results, and code repository links.
5. Ouyang, L., et al. (2022). *Training Language Models to Follow Instructions with Human Feedback.* Advances in Neural Information Processing Systems (NeurIPS), 35. > The foundational RLHF paper, providing the theoretical basis for later verifiable-reward reinforcement learning methods like RLVR.
---
*This article is based on Analemma's official public materials and related frontier research literature. All FARS run data comes from official public disclosures (as of February 23, 2025); paper scores come from the official Agentic Reviewer evaluation reports. Readers can access primary sources at https://analemma.ai/fars and gitlab.com/fars-a.*