A Cursor plugin plus a local swarm that implements the execution-grounded Co-Scientist loop from Schmidgall et al., Accelerating Scientific Research with Gemini in the Real-World (arXiv:2608.26701 / alphaXiv).
It is built for a CS PhD: hypothesis tournaments, Grok swarms, isolated experiments, and papers that are not allowed to exist without logs.
directive → ethics → literature → generate → critique → TrueSkill+UCB
→ evolve → plan → scaffold/full programs → logs
→ manuscript (review − 0.5·plagiarism − 1.0·hallucination)
First-wave AI Scientists optimized an LLM reviewer and fabricated results (Chen 2025; fabrication 80–100% in small audits). Gupta & Pruthi 2025 measured plagiarism up to ~24%. GDM's 2026 extension adds:
- Bayesian idea tournaments (TrueSkill + UCB)
- Evolutionary programs with real execution
- Hallucination clipping against
E_log - A two-layer ethics gate
- A CS result: Agent_H, an inference-time scaffold that beat six frontier models on length-adjusted HealthBench Hard / Professional
This repo turns that paper into something you can run in Cursor with Grok workers.
Repo root is a single Cursor plugin (phd-co-scientist).
- Clone this repository.
- Cursor → Settings → Plugins → add from folder / Git URL.
- Optional: set
XAI_API_KEYin Plugins → Configure (live Grok). Offline mode still runs the full loop on a deterministic proxy.
Slash commands: /co-scientist /sota-survey /ideate /experiment /write-paper /swarm /discover-architecture /phd-weekly
python3 -m pip install -e ".[dev]"
python3 -m pytest tests
python3 -m swarm sota --lab DeepMind
python3 -m swarm ideate "Discover a cheaper tournament scaffold" --generations 4
python3 -m swarm run "Discover a length-calibrated tournament agent on a held-out proxy"
python3 -m swarm swarm "co-scientist SOTA" -n 6Artifacts land in runs/<timestamp>/ (report.json, paper.md, experiment/execution.log).
| Path | What |
|---|---|
skills/ |
15 skills (loop, SOTA, ethics, tournament, paper integrity, Grok swarm, Agent_H search) |
agents/ |
Supervisor + paper coalition (generation, reflection, ranking, evolution, …) |
commands/ |
Slash commands |
rules/ |
Always-on protocol, integrity, safety |
hooks/ |
Fail-open ethics hints on prompts and shell |
swarm/ |
TrueSkill, ethics, reliability, ideation, experiment, paper, CLI |
data/sota_catalog.json |
Curated GDM / Stanford / Sakana / FutureHouse / eval catalog |
docs/ |
Architecture, SOTA, PhD cadence, Agent_H, references |
Paper-diagram filenames (solve.py, factory.py, agents.py, harness.py, readme.txt) sit at repo root and re-export the package.
Full write-up: docs/SOTA.md. Catalog dump: python -m swarm sota.
| Lab | System | Steal |
|---|---|---|
| GDM | Co-Scientist (Nature 2026; 2608.26701) | Tournaments, log-grounded papers, adaptive autonomy |
| GDM | AlphaEvolve | Machine-checkable fitness |
| Stanford | Biomni, Virtual Lab, MLAgentBench | Tool-grounded lab agents; research-task specs |
| Sakana | AI Scientist / v2 | Closed code loop — plus integrity modules they lacked |
| FutureHouse | Robin, PaperQA2 | Typed search / read / data tools |
| OpenAI | HealthBench Hard/Pro, MLE-bench | Held-out rubrics, length-adjusted scores |
| Ancestry | Coscientist, ChemCrow | Hardware/tools as source of truth |
In-silico CS default. No clinical advice. No exploits. No dual-use bio. Ethics hooks fail open if they crash, but the Python gate fails closed on restricted directives.
MIT. The paper this is inspired by is Google's; this repo is an independent implementation of the public architecture for your own research.