OpenNSWM-Lab/FAROS

What it solves

FAROS addresses the lack of grounding and traceability in AI-generated research. It prevents LLMs from generating baseless research ideas by enforcing evidence-based retrieval first, replaces unstructured prose handoffs with structured contracts, and ensures that automated reviews are actionable and linked to real experimental results rather than being generic feedback.

How it works

FAROS operates as a multi-agent system that follows a structured scientific loop:

1. Idea Generation: Uses LLMs (like Qwen) to turn rough interests into specific tasks and methods, grounded in literature retrieval and semantic filtering.
2. PlanPackage: Creates a structured contract containing hypotheses, variables, and acceptance criteria to guide downstream modules.
3. Code & Experiment: Generates executable code in a sandbox, runs experiments, and collects structured metrics and figures.
4. Paper: Drafts a paper linked directly to the original plan and experimental data.
5. ReviewX: An audit system that checks for consistency between claims, evidence, and measurements, feeding results back into the research loop for revision.

Who it’s for

Researchers and AI scientists who need a collaborative system to manage the end-to-end research process—from ideation to publication—while maintaining a strict audit trail of evidence and human oversight.

Highlights

• Evidence-First Approach: Literature must pass quality gates before being used to support a research idea.
• PlanPackage Contract: Uses a typed handoff system so downstream agents don't have to guess the intent of upstream agents.
• ReviewX Closed-Loop: Audits the consistency of claims and evidence, allowing real experimental results to trigger revisions in the research plan.
• Human-in-the-Loop: Integrates human approval and sign-off at critical decision gates to ensure trust and reliability.
• Recoverable Workflows: Supports background tasks and polling to prevent data loss during long-running research tasks.