nanoresearch

nanoresearch

Autonomous ML research in one command. Scout ideas, run experiments, write a paper, survive peer review — overnight.

/nanoresearch "efficient attention mechanisms for long-context LMs"

Wake up to an accepted paper or an honest rejection analysis.

Philosophy

nanoresearch is the irreducible core of ARIS, rebuilt from first principles following Karpathy's autoresearch and the nanorepl philosophy: strip everything that isn't the core algorithm, then layer complexity back one concept at a time.

The algorithm of research is propose → evaluate → keep/discard, applied at three levels:

Level	Phase	What it does
Ideas	Scout	Survey literature, generate ideas, verify novelty, produce experiment spec
Code	Loop	Edit → commit → run → measure → keep or revert (autoresearch)
Writing	Write	Draft section → review → fix (×6 sections), then revise → panel review → fix (×2 passes)
Papers	Review	4 reviewers score → rebuttal → area chair decides → revise if rejected

Pipeline

/nanoresearch "topic"

┌────────┐     ┌─────────────────────┐     ┌──────────────────────────┐     ┌──────────────┐     ┌──────────┐
│ SCOUT  │────▶│        LOOP         │────▶│          WRITE           │────▶│    REVIEW    │────▶│   DONE   │
│        │     │                     │     │                          │     │              │     │          │
│ survey │     │  ┌────────────────┐ │     │  ┌─────────────────────┐ │     │ 4 reviewers  │     │ paper.pdf│
│ ideate │     │  │ edit → commit  │ │     │  │ draft → review → fix│ │     │ + rebuttal   │     │ + reviews│
│ verify │     │  │ → run → measure│ │     │  └──┬────────────┬────┘  │     │ + area chair │     │          │
│ spec   │     │  └──┬──────────┬──┘ │     │  next section  all done  │     │              │     │          │
│        │     │   revert       keep │     │    └──────┬──────┘       │     └──────┬───────┘     └──────────┘
│        │     │     └─────┬────┘    │     │  ┌─────────────────────┐ │            │
│        │     │   budget exceeded?  │     │  │ 4+4 panel → fix     │ │        rejected?
│        │     │   yes ──▶ exit      │     │  └──┬────────────┬─────┘ │            │
└────────┘     └─────────────────────┘     │  next pass    2 passes   │            │
                      ▲                    │     └──────┬─────┘       │            │
                      │                    │  (or memo if weak)       │            │
                      │                    └──────────────────────────┘            │
                      └───────────── REVISE: update spec & resubmit ◀──────────── yes

Setup

# 1. Clone the repo
git clone https://github.com/chrisliu298/nanoresearch.git ~/.nanoresearch

# 2. Register the marketplace — add this to ~/.claude/settings.json:
#    "extraKnownMarketplaces": {
#      "nanoresearch": {
#        "source": { "source": "directory", "path": "~/.nanoresearch" }
#      }
#    }

# 3. Install the plugin
claude plugin install nanoresearch@nanoresearch

# Required: Codex MCP (for GPT-5.4 reviewers and brainstorming)
npm install -g @openai/codex
codex setup  # set model to gpt-5.4
claude mcp add codex -s user -- codex mcp-server

# Required for paper writing: LaTeX
# macOS: brew install --cask mactex && brew install poppler
# Ubuntu: sudo apt install texlive-full latexmk poppler-utils

Usage

# Full pipeline — one command, go to sleep
/nanoresearch "factorized attention for efficient long-context transformers"

# Individual phases
/nanoresearch:scout "topic"                         # just scout
/nanoresearch:loop                                  # just experiment loop
/nanoresearch:write                                 # just paper writing
/nanoresearch:review                                # just peer review

# Override defaults
/nanoresearch "topic" — budget: 8h              # longer experiment budget
/nanoresearch "topic" — loop: 50                # bound to 50 iterations
/nanoresearch "topic" — venue: NeurIPS          # target venue
/nanoresearch "topic" — skip-scout              # skip idea discovery
/nanoresearch "topic" — skip-to-write           # skip to paper writing

Paper Writing

The write phase is a two-sub-phase loop, not a one-shot draft:

Sub-phase 1: Section-by-section drafting. Sections are written in dependency order (method → experiments → related work → introduction → conclusion → abstract). Each section gets a GPT-5.4 xhigh review gate before the next begins — catching issues early before they propagate.