ARIS: AI That Does ML Research While You Sleep
ARIS (Auto-Research-In-Sleep) transforms Claude Code into a complete ML research assistant. From idea discovery through paper writing and rebuttals, this 5.4Kโ GitHub repo automates the entire research pipeline. Wake up to scored papers, fixed weaknesses, run experiments, and rewritten narratives - all autonomous. Works with Claude Code, Codex CLI, Cursor, Trae, and more. Zero dependencies, pure Markdown skills.
ARIS: Let AI Do Your ML Research While You Sleep
Auto-Research-In-Sleep (ARIS) is a revolutionary Claude Code skill collection that automates the entire machine learning research lifecycle. With 5.4K GitHub stars and papers accepted at AAAI 2026, this is production-proven autonomous research.
๐ What ARIS Does (While You Sleep)
- /idea-discovery: Literature survey โ 8-12 ideas โ novelty check โ GPU pilots โ ranked report
- /auto-review-loop: GPT-5.4 reviews paper (5/10โ7.5/10) โ writes experiments โ deploys to GPU โ rewrites
- /paper-writing: Narrative โ claims matrix โ figures โ LaTeX โ submission-ready PDF
- /rebuttal: Parse reviews โ strategy โ safe draft โ GPT stress test โ PASTE_READY.txt
- /research-pipeline: Full end-to-end: idea โ paper โ submit
๐ฏ Real Results
| Paper | Score | Venue | Built With |
|---|---|---|---|
| CS Paper | 8/10 "clear accept" | CS Conference | Claude + GPT-5.4 |
| AAAI Paper | 7/10 "good paper" | AAAI 2026 | Pure Codex CLI |
๐ฅ Key Features
- Cross-model adversarial review: Claude executes, GPT-5.4 critiques (beats self-play)
- GPU automation: Local/remote/Vast.ai rental, auto-debug + retry
- Zero dependencies: Pure Markdown SKILL.md files
- Multi-platform: Claude Code, Codex CLI, Cursor, Trae, Antigravity, OpenClaw
- Free alternatives: GLM-5 + MiniMax-M2.7 (no Claude/OpenAI API needed)
๐ฎ Quick Start (5 Minutes)
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
mkdir -p ~/.claude/skills/
cp -r skills/* ~/.claude/skills/
claude
Then:
/research-pipeline "factorized gap in discrete diffusion LMs"
๐ ๏ธ Workflows
Workflow 1: /idea-discovery "discrete diffusion models" โ literature + 8 ideas + pilots
Workflow 2: /auto-review-loop "my paper topic" โ 4-round improvement loop
Workflow 3: /paper-writing NARRATIVE_REPORT.md โ ICLR/NeurIPS PDF
Workflow 4: /rebuttal "paper/ + reviews" โ venue: ICML
๐ No GPU? No Problem
- Modal serverless GPU:
gpu: modal(community contribution) - Vast.ai rental:
gpu: vastauto-rents/destroys - ModelScope free tier: 2000 calls/day
๐ Why It Works
Single-model self-review = local minima. ARIS uses adversarial collaboration:
Claude Code (fast execution) ร GPT-5.4 xhigh (rigorous critique)
Real score progression (overnight run):
| Round | Score | What Happened |
|---|---|---|
| Initial | 5.0/10 | Borderline reject |
| Round 4 | 7.5/10 โ | 20+ GPU experiments, narrative pivot |
๐ Community Power
- 32 contributors, 12 community skills (proof-writer, paper-poster, grant-proposal)
- Featured: PaperWeekly, awesome-agent-skills, AI Digital Crew
- Adaptations: OpenClaw, Cursor, Trae (ByteDance), Antigravity (Google)
๐ Get Started Today
cd your-research-project
/research-pipeline "your research direction" โ difficulty: nightmare
Wake up to a submission-ready paper. ARIS doesn't replace your insight - it amplifies it 10x.
โญ GitHub Repo | ๐ Full Docs