mirror of
https://github.com/merlinhu1/Battle-tested-Research-Prompts.git
synced 2026-08-26 12:20:55 +02:00
MerlinH
4b069723f7
- OpenAI: Cycle Double Cover conjecture prompt; First Proof solve/verify/refine templates - Anthropic: Riemann zeta 67.2% critical-line prompts + full transcript; protein binder design campaign prompts - Google DeepMind: Co-Scientist agent prompt templates and research goals - Root README index + CONTRIBUTING with firm inclusion criteria (proven yield required)
Co-Scientist — Multi-Agent Hypothesis Generation
| Lab | Google DeepMind / Google Research |
| Model | Gemini 2.0 (multi-agent coalition) |
| Field | Life sciences / biomedical discovery — hypothesis generation |
| Result (yield) | Lab-validated discoveries: novel AML drug-repurposing candidates (KIRA6 confirmed active at clinically relevant concentrations), novel epigenetic anti-fibrotic targets validated in human hepatic organoids (one candidate blocked 91% of a scarring-linked response), and independent re-discovery of an unpublished cf-PICI gene-transfer mechanism. Published in Nature, May 19, 2026, with case studies from Stanford, MIT, Edinburgh, Cambridge, Calico and others. |
| Date | May 19, 2026 (Nature paper) |
| Paper | https://www.nature.com/articles/s41586-026-10644-y (arXiv:2502.18864) |
| Announcement | https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/ |
Why it's battle-tested
Co-Scientist's tournament-of-ideas loop — generate → debate → rank → evolve — with Elo-based ranking and compute-scaled self-improvement, produced hypotheses that survived real laboratory validation across drug repurposing, target discovery, and mechanism elucidation, culminating in a Nature publication.
The prompts
Two layers of prompts drive the system:
- Research goals (human → system): short natural-language objectives
(see
research-goals.md), parsed into a research plan configuration with preferences, attributes, and constraints. - Agent prompts (system → agents): the released per-agent templates for
the Generation, Reflection, Ranking, Evolution, and Meta-review agents
(see
agent-prompts.md).
Notable techniques in the agent prompts:
- Structured role framing ("You are an expert participating in a collaborative discourse…").
- Debate procedures with explicit turn structure, termination conditions, and contribution rules (propose three distinct hypotheses; critically evaluate; conclude with a refined iteration).
- Ranking tournaments with pairwise comparison prompts that disregard numerical scores from prior reviews.
- Meta-review prompts that synthesize recurring critique points into actionable insights.
- Elo-rated outputs correlating with GPQA-diamond accuracy.
Files
agent-prompts.md— Section 9 of the paper (all specialized-agent prompt templates), machine-extracted from the arXiv PDF.research-goals.md— example research goals from published, lab-validated runs.