FAROS (Foundation AutoResearch Operating System) is a multi-agent system for running a research project from an initial question to auditable evidence. It covers topic selection, literature retrieval, research planning, code generation, experiment tracking, paper writing, and peer review, and it records sources, decisions, artifacts, and human feedback at every stage. The current release targets LLM and AI Scientist research.
Instead of chaining loose LLM calls, FAROS starts from literature evidence. Retrieved papers must pass semantic-alignment and quality gates before they can support an idea. Plans are handed between modules as a structured PlanPackage that fixes hypotheses, variables, steps, acceptance criteria, and evidence references, and projects, runs, experiments, figures, papers, and reviews share traceable identifiers. A review module, ReviewX, extracts the paper's claims, checks them against cited evidence and experimental measurements, and emits revision actions that feed back into the next plan. People keep the decisions at plan approval, result interpretation, and final sign-off.
| Stage | Main capabilities | Output |
|---|---|---|
| Idea | Qwen topic coaching, retrieval, semantic filtering, deep reading, novelty and feasibility review | Evidence-supported research candidates |
| PlanPackage | Hypothesis refinement, variables, staged execution, reviewer committee, human approval | Executable and testable research plan |
| Code | Code retrieval, project generation, sandbox execution, static and dynamic evaluation, repair | Runnable experiment project and execution trace |
| Experiment | Metric ingestion, dataset upload, result comparison, publication figures | Structured results and reusable figures |
| Paper | Brief, outline, section drafting, citations, LaTeX and PDF generation | Paper project linked to plans and experiments |
| ReviewX | Claim extraction, evidence alignment, consistency and reliability checks, human feedback | Auditable review report and revision plan |
The runtime follows a Blueprint, Capability, Profile, and Provider design so new workflows and domains can be added without changing the orchestration core. It has a FastAPI backend and a React workbench with a bilingual Chinese and English interface. Qwen is the recommended model provider, and provider credentials are configured per user and stored encrypted. The project describes itself as a release candidate for competition validation and research prototypes, not an unsupervised replacement for researchers.
Features
- Evidence-gated ideation: literature retrieval, semantic alignment, and deep reading run before ideas are generated
- Topic coaching: Qwen turns a rough research interest into a task, method, dataset, and evaluation target
- PlanPackage contract: a typed handoff carrying hypotheses, variables, steps, acceptance criteria, and evidence references
- Code and experiments: project generation, sandboxed execution, automatic repair, metric ingestion, and publication figures
- Paper drafting: briefs, outlines, section writing, citation management, and LaTeX and PDF output
- ReviewX: claim-evidence-measurement consistency checks that produce actionable revision plans
- Human gates: people approve plans, interpret results, and sign off, with all feedback versioned
- Recoverable jobs: long-running tasks run in the background and survive browser disconnects without duplicate model calls
- Extensible runtime: Blueprint, Capability, Profile, and Provider components for new workflows and domains
- Team isolation: per-user provider configuration with encrypted API key storage and a bilingual UI
