FARS: Deep Research Report on a Fully Automated Research System

1. System Overview and Core Breakthroughs

1.1 FARS System Definition

Fully Automated Research System (FARS) is an end-to-end, AI-driven, multi-agent research platform officially released by Analemma on February 12, 2026. It targets full-process automation of scientific research—from hypothesis generation through experimental design and execution to academic paper writing—without direct human intervention.

#### Developer: Analemma

Analemma was founded in March 2025 by an academic team largely from Fudan University's MOSS large language model group and Shanghai AI Lab's InternLM team. Founder Dr. Tianxiang Sun (PhD, Computer Science, Fudan; advised by Prof. Xuanjing Huang and Prof. Xipeng Qiu) was a core developer of MOSS, China's first open-source conversational LLM, with over 4,200 Google Scholar citations. The company has closed an angel round (tens of millions USD) led by prominent investors including Gao Rong, Sequoia China, Light Source Capital, MiraclePlus, Meituan Longzhu, Jiacheng Capital, and Shanghai Future Industry Fund.

Analemma also offers Lemma (lemma.analemma.ai), an "assisted driving" productivity tool for literature review, deep research, and code experiments, while FARS targets "autonomous driving"—fully autonomous, end-to-end automation.

#### Core Mission

FARS's mission is "research industrialization": transforming a traditionally human-centered craft into a scalable, reproducible automation pipeline with three progressive layers—process automation, continuous 24/7 operation, and quality assurance.

1.2 The "FARS-100" Public Livestream Experiment

#### Runtime

Launched February 13, 2026, FARS-100 ran continuously for 228 hours, 28 minutes, 33 seconds (≈9.5 days) with zero human intervention. The livestream at https://analemma.ai/fars displayed real-time project queues and outputs.

#### Output

• 244 research hypotheses generated
• 100 short papers completed
• ≈41% hypothesis-to-paper conversion

The system deliberately favors short papers that report single, bounded contributions—including negative results—to counter publication bias.

#### Production Efficiency

| Dimension | FARS | Human Researcher | Magnitude Gap | |---|---|---|---| | Per-paper time | ~2.28 hours | 3–6 months (90–180 days) | ~950–1900× | | Daily output | ~10.5 papers | ~0.005–0.01 papers | ~1000–2000× | | Continuous operation | 24/7 | Limited by biology | Theoretically unlimited |

#### Resource Consumption

| Resource | Consumption | Notes | |---|---|---| | LLM tokens | 11.4 billion | Multi-vendor mix | | GPU cluster | 160 NVIDIA GPUs | Encapsulated as Experiment tools | | Total operating cost | ~$104,000 | 9.5-day cycle | | Per-paper cost | ~$1,040 | Including token and compute depreciation |

#### Transparency Mechanisms

• Real-time livestream of system state, queues, and outputs
• arXiv publication of all papers with explicit AI-generation tagging
• GitHub code release under the account github.com/fars-analemma

The FARS source code itself is not open-sourced, which has sparked discussion about the balance between open science and commercial interests.

1.3 Historical Significance

#### First Verifiable End-to-End Automation

FARS-100 is the first public, verifiable demonstration that a hypothesis-driven, end-to-end automated research pipeline is engineering-feasible. It validated multi-agent orchestration for long-chain tasks, the hypothesis-driven paradigm, and shared-filesystem workspace design.

#### 24/7 "Research Factory" Model

It established the Research Factory model: standardized Ideation→Planning→Experiment→Writing pipelines, swappable components, internal quality control, and linearly scalable capacity via additional compute.

#### AI4AI at Scale

FARS-100 marks AI4AI's transition from concept validation to scaled deployment, joining precedents such as DeepMind's AlphaEvolve and Sakana AI's AI Scientist v2.

---

2. Technical Architecture and Operating Principles

2.1 Multi-Agent System Architecture

#### Design Philosophy: First Principles

FARS derives its architecture from first principles—analyzing research into a minimal unit of "hypothesis + verification"—and is optimized for efficient, reliable expansion of the knowledge frontier. Output is short, single-contribution papers that welcome negative results.

#### Four Core Features

| Feature | Implementation | Value | |---|---|---| | Modular specialization | Ideation/Planning/Experiment/Writing | Lower complexity, independent upgrades | | Asynchronous collaboration | Shared filesystem | Loose coupling, fault tolerance | | Pipeline scheduling | Project queues across stages | Resource maximization, predictable throughput | | Closed-loop iteration | Auto-rollback on failure | Self-correction, continuous optimization |

#### Infrastructure

• Shared filesystem: workspace plus persistent memory for inter-agent state
• 160-GPU unified dispatch interface: encapsulates training and inference tools for the Experiment agent
• Multi-vendor LLM gateway: dynamic selection across GPT-4, Claude, Gemini, Llama, etc.

2.2 Four Core Agents

#### Ideation Agent Literature scanning, knowledge integration, and hypothesis generation across nine predefined research directions (e.g., RLVR, model architecture innovation, diffusion language models). Balances exploration vs. exploitation and auto-evaluates novelty, feasibility, and value.

#### Planning Agent Converts hypotheses into executable experimental plans: model architecture choices, datasets and metrics, training strategies and hyperparameter search spaces, controls/ablations, and resource estimates. Uses chain-of-thought-style reasoning.

#### Experiment Agent Translates plans into PyTorch/JAX code, trains and runs experiments on the GPU cluster, handles errors, performs basic analysis, and writes results back to the filesystem. Supports dynamic iteration based on intermediate results.

#### Writing Agent Assembles all prior artifacts into a structured academic short paper (introduction, methods, experiments, conclusions), generates figures, and manages references. Outputs are reviewed by at least three senior researchers before arXiv submission.

2.3 Coordination and Workflow

#### Communication

Shared-filesystem asynchronous messaging: each agent writes standardized artifacts (JSON, Markdown, Python) to designated paths; downstream agents poll or watch for changes.

#### Scheduling

Pipeline-style serial processing with parallel project queues. Includes priority mechanisms, load balancing, resource reservation, and timeout handling.

#### Iteration

Auto-rollback on failure: experiment errors trigger local debugging, replanning, or hypothesis replacement.

#### Quality Control

Human review gate: at least three senior researchers review each paper before arXiv release, covering factual accuracy, logical coherence, academic integrity, and proper AI-generation labeling.

2.4 Technical Limits and Boundary Conditions

• Domain: only validated for AI/LLM research (AI4AI)
• Compute dependency: requires 160-GPU-class cluster; hard to reproduce by individuals/small teams
• Experiment types: cannot run experiments requiring human participants
• Scale ceiling: does not currently support large-scale pre-training

---

3. Output Quality Assessment

3.1 Evaluation Methodology

#### Tool: Stanford's Agentic Reviewer (paperreview.ai)

An LLM-based reviewer trained/prompt-engineered for academic peer review. Uses ICLR (International Conference on Learning Representations) criteria—Technical Correctness, Novelty, Significance, Clarity, Relevance.

#### Tool Reliability

| Comparison | Spearman Correlation | Interpretation | |---|---|---| | Human vs. human | 0.41 | Baseline inter-rater agreement | | AI vs. human | 0.42 | Matches human-level consistency |

3.2 Quantitative Results

| Statistic | Value | |---|---| | Mean score | 5.05 | | Range | 3.0–6.3 | | Distribution | Concentrated near 5.0 | | High-scoring (>6.0) | Very few |

#### Comparison with Human Submissions

| Group | Mean Score | |---|---| | FARS output | 5.05 | | ICLR 2026 submissions (average) | 4.21 | | ICLR 2026 accepted papers (average) | 5.39 |

FARS sits clearly above the submission average but below the acceptance threshold—a "stable mid-band output machine."

3.3 Qualitative Case Studies

#### Success: FA0042 — Text Embedding Optimization

Addresses the trade-off between bidirectional (high quality, KV-cache-inefficient) and causal (efficient, weaker representation) attention. FARS proposed a Train–Inference Separation approach: train bidirectionally, switch to causal at inference, bridged by a Gated Gumbel-Softmax Mixture (GG-SM). On LoCoV1 long-document retrieval, the student model reached NDCG@10 = 0.284, beating all baselines including the teacher's 0.212. The GG-SM technique from Ant Group was integrated within 3 days of release.

#### Failure: FA0121 — "Algorithmic Honesty"

Targeted the "hot/cold" gating bias in DeepSeek's Engram sparse architecture. Proposed Counterfactual Gate Supervision (CGS): compute loss differences between fully-open and fully-closed gating to estimate n-gram utility. The result was largely negative—CGS underperformed simply training longer. FARS honestly reported the negative result and analyzed the coupled training dynamics. This exemplifies "algorithmic honesty" addressing the File Drawer Problem.

3.4 Evaluation Caveats

• Output is short papers; direct comparability with long-form submissions is imperfect
• System was not optimized for any specific venue
• AI review scores are reference points, not definitive judgments
• A complementary human expert review is ongoing; a comprehensive quality report is forthcoming

---

4. Impact Analysis for Researchers

4.1 Direct Shocks and Sources of Anxiety

• Efficiency gap: 2.17 hours per paper vs. 3–6 months for humans
• Cost restructuring: ~$1,040/paper vs. $10,000–$80,000 for human-led research
• Scarcity collapse: industrialization undermines the traditional "academic currency" function of papers
• Identity crisis: ideation, experimentation, and writing—the three core research skills—are systematically replicated

4.2 Irreplaceable Human Value

• Critical thinking and originality: AI recombines statistical patterns rather than generating paradigm shifts
• Cross-domain intuition and metaphorical transfer: structural similarity perception across disparate fields
• Value judgments about research direction: deciding what problems matter most
• Complex physical-world experiments and human-subject research: beyond current AI reach
• Academic community and knowledge传承: mentorship, scholarly networks, disciplinary culture

4.3 Role Transformation

| From | Toward | Capabilities Required | |---|---|---| | Paper producer | Research architect | Problem-space mapping, hypothesis curation, system supervision | | Solo researcher | Human–AI collaboration manager | Multi-agent orchestration, compute strategy, cross-system integration | | Technical executor | Academic value gatekeeper | Real-value assessment, new evaluation criteria, integrity enforcement |

---

5. Systematic Reflections on Future Research Models

5.1 Evolution of Research Paradigms

• Current stage: AI4AI—domain-limited validation
• Near-term expansion: physical simulation–friendly domains (computational physics, chemistry, materials science)
• Mid-term vision: hybrid human–AI research teams as the norm
• Long-term vision: AI systems "hiring" humans via crowdsourcing/remote collaboration when physical or human-subject experiments are unavoidable—reversing traditional human–machine collaboration boundaries

5.2 Key Strategic Considerations

Effective integration requires designing collaboration interfaces, incentive structures, and governance frameworks that enable human researchers to benefit from AI collaboration rather than perceive it as competition. The system is best understood as a tool that shifts researcher focus from mechanical execution toward strategic direction-setting, value judgment, and the cultivation of disciplinary culture.

---

Key Points

• FARS-100 ran autonomously for 228h 28m 33s, producing 100 short papers from 244 hypotheses at ~2.17 hours per paper
• Total cost ≈ $104,000 across 11.4 billion tokens and a 160-GPU cluster, implying ~$1,040 per paper
• Four-agent pipeline (Ideation→Planning→Experiment→Writing) coordinated via shared filesystem
• Agentic Reviewer mean score 5.05 vs. 4.21 submission average and 5.39 acceptance threshold on ICLR criteria
• FA0042 (text embedding, GG-SM integration in 3 days) and FA0121 (negative result, "algorithmic honesty") illustrate both capability and integrity
• Researcher roles shift from paper producer to research architect, hybrid team manager, and academic value gatekeeper
• Domain currently limited to AI4AI; expansion toward computational physics/chemistry/materials is most feasible short-term