FARS: Deep-Dive Report on a Fully Automated Research System That Ran for 228 Hours

Forum topic · ✨步子哥 · 2026-02-25

Summary

FARS (Fully Automated Research System) is an end-to-end, AI-driven multi-agent research pipeline developed by Analemma, an AI startup founded by former Fudan University MOSS and Shanghai AI Lab InternLM researchers. Launched in February 2026, FARS automates the full research lifecycle—hypothesis generation, experiment planning, execution, and paper writing—through four specialized agents (Ideation, Planning, Experiment, Writing) coordinating via a shared file system and a 160-GPU cluster. In its flagship FARS-100 live-streamed experiment, the system ran unattended for 228 hours 28 minutes 33 seconds, generated 244 research hypotheses, and completed 100 short papers at an average pace of roughly one paper every 2.3 hours, consuming 11.4 billion tokens at a total cost of about $104,000 (roughly $1,040 per paper). Papers were published on arXiv with code on GitHub, though each release passed review by at least three human researchers. Scored by Stanford's Agentic Reviewer under ICLR criteria, the papers averaged 5.05—above the average human submission (4.21) but below the acceptance threshold (5.39). The report analyzes FARS's architecture, output quality, limitations (AI/LLM research only, heavy compute dependence), and its implications for researchers' evolving roles.

Overview

FARS (Fully Automated Research System) is an end-to-end, multi-agent AI research system released by Analemma on February 12, 2026. Unlike AI research *assistants*, FARS is positioned as a self-running "autopilot" research pipeline—automating literature review, hypothesis generation, experiment design, code implementation, analysis, and paper writing. Analemma was founded in March 2025 by Sun Tianxiang (lead developer of MOSS, Fudan University) with a team from the MOSS and InternLM projects, and has raised tens of millions of USD in angel funding from investors including Hongshan (Sequoia China), Gaorong, and Meituan Longzhu. The company also offers Lemma, an "assisted-driving" research productivity tool, while FARS targets full autonomy.

Key points

The FARS-100 live experiment

Ran continuously for 228h 28m 33s (~9.5 days) during Chinese New Year 2026, fully unattended and live-streamed at https://analemma.ai/fars.
Generated 244 hypotheses and completed 100 short papers (~41% conversion rate); average of one paper per ~2.3 hours, vs. 3–6 months per paper for human researchers.
Consumed 11.4 billion tokens and used a 160× NVIDIA GPU cluster; total cost ≈ $104,000 (~$1,040 per paper).
Papers were published on arXiv (explicitly AI-labeled) and code committed to github.com/fars-analemma; however, FARS's own source code is not open.

Architecture

Four specialized agents—Ideation, Planning, Experiment, Writing—collaborate asynchronously via a shared file system (workspace + persistent memory) with pipeline scheduling and parallel project queues.
Failed experiments trigger automatic rollback to Planning or Ideation stages.
Every paper passes review by at least 3 senior human researchers before arXiv submission, a caveat to the "unattended" claim.

Quality assessment

Evaluated with Stanford's Agentic Reviewer (paperreview.ai) under ICLR criteria; the tool's agreement with humans (Spearman 0.42) matches human-human agreement (0.41).
FARS papers averaged 5.05 (range 3.0–6.3)—above ICLR 2026 average submissions (4.21) but below accepted-paper average (5.39).
Case study FA0042 (text embeddings, train-bidirectional / inference-causal with GG-SM transition) showed rapid 3-day adoption of a new external technique; case study FA0121 honestly reported a failed counterfactual gate supervision method on DeepSeek's Engram architecture.
Short-paper format and non-optimization for specific venues limit direct comparisons; a human expert review is ongoing.

Limitations

Restricted to AI/LLM research (AI4AI); cannot run physical experiments, human-subject studies, expert evaluation, or large-scale pretraining.
Heavy compute dependence makes replication difficult for individuals or small labs; the system is closed-source.

Implications for researchers

Efficiency and cost disruption challenge paper-count-based careers; scarcity of publication as an academic currency erodes.
Humans retain advantages in paradigm-shifting originality, cross-domain intuition, value judgments on what matters, physical-world experiments, and academic community building/mentorship.
Recommended role shifts: from paper producer to research architect, from solo researcher to human-AI collaboration manager, from executor to academic value gatekeeper.
Future trajectory: near-term expansion into simulation-heavy fields (computational physics/chemistry, materials science); mid-term hybrid human-AI research teams; longer-term scenarios where AI "hires" humans for experiments it cannot perform.