aris-autonomous-ml-research

---
name: aris-autonomous-ml-research
description: Autonomous ML research workflows using ARIS (Auto-Research-In-Sleep) — Markdown-only skills for cross-model paper review, idea discovery, experiment automation, and paper writing with Claude Code, Codex, or any LLM agent.
triggers:
  - "set up ARIS for autonomous research"
  - "run research pipeline while I sleep"
  - "automate ML paper writing with Claude Code"
  - "cross-model review loop for my paper"
  - "use ARIS to find research ideas"
  - "run experiment automation with ARIS"
  - "set up auto paper review workflow"
  - "write rebuttal with ARIS"
---

# ARIS — Auto-Research-In-Sleep

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

ARIS is a **zero-dependency, Markdown-only** autonomous ML research system. Every "skill" is a plain `SKILL.md` file that any LLM agent can read and execute. It orchestrates cross-model collaboration — one model executes research (Claude Code, Codex, etc.) while another acts as adversarial reviewer (GPT-5.4, Gemini, GLM, MiniMax, etc.) to break self-play blind spots.

**Core value**: going from research direction → paper ideas → experiments → written paper → rebuttal, autonomously, overnight.

---

## Installation

### 1. Clone the Repository

```bash
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
cd Auto-claude-code-research-in-sleep

No pip install, no Docker, no daemon. The entire system is Markdown files.

2. Install Claude Code (Primary Agent)

npm install -g @anthropic-ai/claude-code

3. Install Codex MCP (Cross-Model Reviewer)

npm install -g @openai/codex

Configure Claude Code to use the Codex MCP server by adding to your Claude Code config (~/.claude/settings.json):

{
  "mcpServers": {
    "codex": {
      "command": "codex",
      "args": ["mcp"],
      "env": {
        "OPENAI_API_KEY": "$OPENAI_API_KEY"
      }
    }
  }
}

4. Copy Skills into Claude Code

# Copy all skills to Claude Code's custom skills directory
cp -r skills/claude-code/ ~/.claude/skills/

# Or symlink to stay up to date
ln -s $(pwd)/skills/claude-code ~/.claude/skills/aris

5. Set Environment Variables

# Required for Claude Code
export ANTHROPIC_API_KEY=your_anthropic_key

# Required for cross-model review (GPT-5.4 as reviewer)
export OPENAI_API_KEY=your_openai_key

# Optional: alternative reviewer models (no OpenAI needed)
export LLM_REVIEWER_BASE_URL=https://api.minimax.chat/v1
export LLM_REVIEWER_API_KEY=your_minimax_key
export LLM_REVIEWER_MODEL=MiniMax-M2.7

---

Alternative Model Combinations (No Claude/OpenAI Required)

ARIS works with any OpenAI-compatible API. Configure the llm-chat MCP server:

{
  "mcpServers": {
    "llm-chat": {
      "command": "node",
      "args": ["mcp-servers/llm-chat/index.js"],
      "env": {
        "LLM_BASE_URL": "$LLM_REVIEWER_BASE_URL",
        "LLM_API_KEY": "$LLM_REVIEWER_API_KEY",
        "LLM_MODEL": "$LLM_REVIEWER_MODEL"
      }
    }
  }
}

Tested combinations:

Executor	Reviewer	Config
Claude Code	GPT-5.4 xhigh	Default
Codex CLI	Gemini	Guide
Claude Code	MiniMax-M2.7	LLM_BASE_URL=https://api.minimax.chat/v1
Claude Code	GLM-5	LLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
MiniMax-M2.7	GLM-5	Guide
Codex CLI	Claude	Swap executor/reviewer

---

Core Workflows

Workflow 0: Full Pipeline (Start Here)

/research-pipeline "factorized gap in discrete diffusion LMs"

With a reference paper and base repo:

/research-pipeline "improve method X" — ref paper: https://arxiv.org/abs/2406.04329, base repo: https://github.com/org/project

ARIS will:

1. Read the paper → find weaknesses
2. Clone the codebase
3. Generate ideas that fix those weaknesses using that code
4. Run experiments
5. Write the paper

Parameters:

/research-pipeline "topic"
  — ref paper: <arxiv_url>       # Optional: paper to improve
  — base repo: <github_url>      # Optional: codebase to build on
  — venue: ICML                  # Target venue (default: ICML)
  — compact: true                # Lean summaries for short-context models

---

Workflow 1: Idea Discovery

/idea-discovery "discrete diffusion language models"

Scans literature, identifies gaps, generates novel research directions, scores each idea for novelty/feasibility, and outputs a ranked proposal list.

---

Workflow 1.5: Experiment Bridge

/experiment-bridge "run ablation on temperature scaling" — code review: true

Cross-model code review before GPU deployment (enabled by default). Catches bugs, confirms experimental validity, then runs.

# Example: what experiment-bridge automates
# 1. Claude Code writes training script
# 2. GPT-5.4 reviews the code (code review gate)
# 3. If approved → submits to GPU cluster
# 4. Monitors via W&B API
import wandb

api = wandb.Api()
runs = api.runs("your-entity/your-project")
for run in runs:
    print(run.name, run.summary.get("val_loss", None))

---

Workflow 2: Paper Writing

/paper-writing "results/" — venue: NeurIPS

Generates LaTeX paper from experiment results. Anti-hallucination enforced: every citation verified via DBLP → CrossRef → [VERIFY] tag if unconfirmed.

Venue templates available: ICML, NeurIPS, ICLR, CVPR, ACL, AAAI, ACM MM

---

Workflow 3: Auto Review Loop

/auto-review "paper.pdf"

The core ARIS loop:

1. Claude Code reads the paper
2. GPT-5.4 reviews as adversarial critic
3. Claude Code rewrites based on critique
4. Score tracked across rounds (target: 8/10 "clear accept")
5. Loop repeats until convergence or max rounds

Score progression: 5.2 → 6.1 → 7.3 → 8.0 ✓

---

Workflow 4: Rebuttal

/rebuttal "paper/ + reviews" — venue: ICML, character limit: 5000

Parameters:

Parameter	Default	Description
venue	ICML	Target venue
character limit	required	Hard limit for submission
quick mode	false	Stop after parsing + strategy (no draft)
auto experiment	false	Auto-run supplementary experiments
max stress test rounds	1	GPT-5.4 stress-test iterations
max followup rounds	3	Per-reviewer follow-up limit

Three safety gates (rebuttal won't finalize if any fails):

• 🔒 No fabrication — every claim maps to paper/review/user-confirmed result
• 🔒 No overpromise — every promise is user-approved
• 🔒 Full coverage — every reviewer concern is tracked

Outputs:

• PASTE_READY.txt — exact char count, paste directly to venue
• REBUTTAL_DRAFT_rich.md — extended version for manual editing

---

Bonus: Slides and Poster

# Conference presentation
/paper-slides "paper/"     # → Beamer PDF + PPTX + speaker notes + Q&A prep

# Conference poster
/paper-poster "paper/"     # → A0/A1 poster PDF + editable PPTX + SVG

---

Standalone Skills

These skills can be invoked independently or are integrated into the core workflows:

Skill	Command	Description
Research Refine	/research-refine	Turn vague ideas into anchored proposals
Experiment Plan	/experiment-plan	Claim-driven experiment roadmaps
Training Check	/training-check	Validate training runs before full launch
Result to Claim	/result-to-claim	Convert raw results to paper claims
Ablation Planner	/ablation-planner	Design ablation study structure
Formula Derivation	/formula-derivation	Research formula development and verification
Grant Proposal	/grant-proposal	Write grant proposals from research
Paper Illustration	/paper-illustration	Generate figures (Gemini-powered)
Citation Claw	/citation-claw	Verify and format citations

---

Session Recovery & Compact Mode

For short-context models or after interruption:

/research-pipeline "topic" — compact: true

Generates lean summary files at each checkpoint. Resume after interruption:

/research-refine — resume: true

ARIS auto-checkpoints the research-refine workflow and resumes from last completed phase.

---

Codex CLI Native Skills

Full skill set available for OpenAI Codex without Claude Code:

cd skills/skills-codex/
codex "run idea-discovery on discrete diffusion"

---

MCP Server: llm-chat

The llm-chat MCP server bridges any OpenAI-compatible API as a reviewer. Start it manually for debugging:

cd mcp-servers/llm-chat/
node index.js

Environment variables:

export LLM_BASE_URL=https://api.openai.com/v1   # Any OpenAI-compatible endpoint
export LLM_API_KEY=$OPENAI_API_KEY
export LLM_MODEL=gpt-4o                          # Any model name

---

Free Tier via ModelScope

Zero-cost option — no API key required:

# See full guide: docs/MODELSCOPE_GUIDE.md
export MODELSCOPE_API_KEY=your_modelscope_token
export LLM_BASE_URL=https://api-inference.modelscope.cn/v1
export LLM_MODEL=Qwen/Qwen2.5-72B-Instruct

---

Input Templates

Templates for every workflow live in templates/:

ls templates/
# idea-discovery.md
# experiment-bridge.md
# paper-writing.md
# auto-review.md
# rebuttal.md
# research-refine.md

Use them to structure your inputs:

cat templates/rebuttal.md
# Fill in: paper path, review text, venue, character limit
# Then: /rebuttal [filled template]

---

Directory Structure

Auto-claude-code-research-in-sleep/
├── skills/
│   ├── claude-code/          # Claude Code SKILL.md files
│   ├── skills-codex/         # Codex CLI native skills
│   ├── idea-discovery/
│   ├── experiment-bridge/
│   ├── paper-writing/
│   ├── auto-review/
│   ├── rebuttal/             SKILL.md  ← each is a single readable file
│   ├── paper-slides/
│   ├── paper-poster/
│   ├── research-refine/
│   ├── formula-derivation/
│   └── ...
├── mcp-servers/
│   └── llm-chat/             # Universal reviewer bridge
├── templates/                # Input templates for every workflow
├── docs/
│   ├── CURSOR_ADAPTATION.md
│   ├── TRAE_ARIS_RUNBOOK_EN.md
│   ├── ANTIGRAVITY_ADAPTATION.md
│   ├── MODELSCOPE_GUIDE.md
│   ├── MiniMax-GLM-Configuration.md
│   └── CODEX_GEMINI_REVIEW_GUIDE.md
└── README.md

---

Troubleshooting

Cross-model review not triggering:

• Check MCP server is running: codex mcp or node mcp-servers/llm-chat/index.js
• Verify OPENAI_API_KEY or LLM_API_KEY is set
• Check Claude Code MCP config in ~/.claude/settings.json

W&B metrics not loading:

import wandb
# Ensure you're logged in
wandb.login(key=os.environ["WANDB_API_KEY"])
api = wandb.Api()
# Use full entity/project path
runs = api.runs("your-entity/your-project")

Context window exceeded mid-workflow:

/research-pipeline "topic" — compact: true

Then resume with — resume: true on the next interrupted skill.

Citation hallucination warnings ([VERIFY] tags): These are intentional — ARIS flags unverified citations rather than silently hallucinating. Manually verify flagged citations before submission.

Rebuttal exceeds character limit: Increase max stress test rounds — each round trims the draft:

/rebuttal "paper/ + reviews" — character limit: 5000, max stress test rounds: 3

ModelScope free tier rate limits: Add delay between skill calls or switch to a paid endpoint for overnight runs.

---

Why Two Models (Not One, Not Four)

• 1 model self-reviewing → local minima, blind spots (stochastic bandit)
• 2 models cross-reviewing → adversarial critique breaks blind spots (adversarial bandit)
• 4+ models → diminishing returns, 2-4× API cost, coordination overhead

Claude Code = fast fluid execution. GPT-5.4/Gemini/GLM = slower, more deliberate critique. Speed × Rigor = better outcomes than either model alone.

---

Community & Citation

@software{aris2026,
  title  = {ARIS: Auto-Research-In-Sleep},
  author = {wanshuiyin},
  year   = {2026},
  url    = {https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep}
}

Join the community: GitHub Discussions

Papers accepted using ARIS: CS Conference (8/10 "clear accept"), AAAI 2026 Main Technical (7/10 "good paper, accept").