During the 2026 Chinese New Year holiday, while most of the country celebrated, Shanghai-based AI company Analemma ran an unusual livestream: FARS (Fully Automated Research System), an end-to-end AI research pipeline that selects its own topics, runs its own experiments, and writes its own papers. The system has run continuously for over 270 hours and produced more than 100 academic papers—about one every two hours.
- Livestream: https://analemma.ai/fars
- Papers and code (open-sourced in real time): https://github.com/fars-analemma
- A compute cluster of 160 NVIDIA GPUs
- API access to most mainstream large models (GPT, Llama, Qwen, etc.)
- A shared file system acting as working memory
- Average score: 5.05 (out of 10), range 3.0–6.3
- Human ICLR 2026 submissions average: 4.21
- ICLR 2026 accepted papers average: 5.39
- Paradigm-shifting innovation — it excels at incremental work within existing frameworks; breakthroughs like Transformer required human intuition.
- Deep interdisciplinary integration — real breakthroughs often occur at intersections of fields.
- Value judgments — deciding what to study, publish, or shelve requires human oversight.
- Analemma blog: https://analemma.ai/blog/introducing-fars/
- Livestream: https://analemma.ai/fars
- GitHub: https://github.com/fars-analemma
How FARS Works
FARS is not a single model but a multi-agent collaboration system with four specialized "AI researchers":
| Agent | Role | Human Equivalent | |-------|------|------------------| | Ideation | Reads literature, finds gaps, proposes hypotheses | Postdoc/researcher | | Planning | Designs experiments | Experimental designer | | Experiment | Writes code, runs experiments, analyzes data | Algorithm engineer | | Writing | Drafts papers | Academic writer |
The agents coordinate through a shared file system: Ideation proposes hypotheses → Planning designs the study → Experiment executes → Writing integrates results into a paper, with no human intervention. Supporting infrastructure includes:
Quality Assessment
The team blind-reviewed the first 100 papers using a Stanford agentic reviewer simulating ICLR review standards:
In short: FARS's average output exceeds the average human submission but falls just short of the acceptance bar.
Notable characteristics of the papers:
1. Single focused contribution — each paper solves one specific problem, in a concise format. 2. Negative results reported honestly — falsified hypotheses are published as-is, a rarity in academia. 3. Frontier topics — RL-based post-training of language models, efficiency of diffusion language models, long-term memory for AI agents, test-time compute scaling, and non-Transformer architectures.
Cost and Efficiency
| Metric | FARS | Traditional Human Research | |--------|------|----------------------------| | Output speed | ~2 hours/paper | 3–6 months/paper | | Cost per paper | ~$1,000 | Hard to estimate (labor + equipment + time) | | Token consumption | ~114M/paper | N/A | | Personnel | 0 (unattended) | 1–5 full-time |
Over 9.5 days, FARS produced 100 papers for roughly $100,000 total. The comparison isn't perfectly fair—FARS's papers are short exploratory work, while doctoral research is deep and systematic—but it demonstrates that research pipelines are viable at scale.
The Team
FARS was built by Analemma (Shanghai), whose core team comes from Fudan University's MOSS project. Founder Sun Tianxiang, a Fudan PhD under professors Huang Xuanjing and Qiu Xipeng, was a lead developer of MOSS, China's first open-source conversational LLM, with 4,200+ Google Scholar citations. The ~15-person team (average age under 30) also includes members of the InternLM project. Analemma's other product, Lemma, is an AI research-assistant tool; FARS represents the "full self-driving" version of research.
Controversies and Limitations
1. Domain limits: FARS only does "AI researching AI" (AI4AI); it cannot handle physical experiments (biology, chemistry, materials) or studies requiring human subjects. 2. Compute barrier: 160 GPUs put it out of reach for most labs—it's a demonstration of a research factory, not a tool for everyone. 3. Quality control: each paper is human-reviewed by three senior researchers before arXiv upload, but the overall impact of 100 papers remains unproven. 4. Research ethics: whether mass AI-generated papers will add noise to academic publishing, and how to ensure integrity, are unresolved questions.
What This Means for Human Scientists
FARS raises the question of where human researchers fit when AI can produce papers around the clock. What it cannot (yet) do:
The emerging division of labor: humans propose good questions and provide insight, ethics, and depth; AI handles literature scanning, hypothesis candidates, experiment execution, and drafting. Future research may look like: humans ask the questions, AI explores and validates, humans then think deeply about the findings.
Conclusion
FARS's 270-hour livestream is a milestone for automated research, proving that an end-to-end pipeline can run sustainably and produce competitively rated work. It is still in a "trading compute for intelligence" phase—over 100 million tokens and ~$1,000 per paper—and far from low-cost, large-scale, high-quality output. Its larger significance may be freeing human scientists from repetitive labor to focus on what matters: curiosity-driven exploration and expanding the boundaries of human knowledge.