A tri-level co-evolving multi-agent research automation system.
Re-implementation of arXiv:2605.10813 with a ChatGPT-style web UI and field-agnostic prompts.
GitHub • Original paper • Quickstart
What is NanoResearch?
NanoResearch takes a one-line research idea — in any scholarly field — and rides it through ideation, planning, experimentation, analysis, writing, and review to produce a downloadable LaTeX paper, while learning your preferences so the next run feels more like you.
It is built around three persistent, per-user stores that co-evolve over time:
| Store | What it holds | Updated when |
|---|---|---|
Skill Bank 𝒮 |
Reusable procedural rules (“design one-factor ablations”) | Every stage end (distillation) |
Memory Module ℳ |
Project-specific facts (“for PubMedQA, baselines are X/Y/Z”) | Every stage end (distillation) |
Planner LoRA π_θ |
A per-user adapter on Qwen2.5-7B | Every feedback round (SDPO, Eq. 14-15) |
Pipeline
