// the find
OpenNSWM-Lab/FAROS
A blueprint-driven AutoResearch runtime for orchestrating AI research workflows from idea generation and experiments to paper writing and peer review.
FAROS is a Python framework for orchestrating AI research pipelines — idea generation, experiment scaffolding, LaTeX paper drafting, and simulated peer review — through a blueprint/capability/provider abstraction layer. It targets ML researchers who want to automate the boilerplate of writing and reviewing papers. The architecture is more principled than most 'AI scientist' repos, but the current release is mostly scaffolding around a single real workflow.
The blueprint/capability/profile/provider layering is a genuine design decision, not just naming things differently — it means you could swap LLM providers or add a human-in-the-loop step without rewriting the orchestrator. Venue-aware LaTeX generation with real ICML/NeurIPS/ICLR/ACL templates is practical and saves real time. The README is unusually honest about what's missing, including a detailed TODO list that names the actual gaps rather than hiding them. Plan-mode (dry run that builds the execution plan without firing LLM calls) is a useful first-class feature for iterating on blueprints cheaply.
The experiment stage is a scaffold, not an executor — it provisions a code workspace but doesn't run code or collect results, which means the paper being written has no real numbers behind it. This is the core problem and it's unresolved. File-backed persistence for runs, events, and artifacts means no safe concurrent access, no crash recovery mid-run, and no useful querying — this will bite anyone running multiple long workflows. Despite the elaborate registry and loader infrastructure, there is exactly one blueprint and one real profile; the abstraction is buying extensibility that hasn't been needed yet and adds friction to reading the code. Reviewer simulation is structurally just another LLM call with no literature retrieval grounding, so it can only critique what the paper itself says — it won't catch claims that contradict published work.