Skip to main content
Back

ARIS Lets Claude and GPT-5.5 Argue Over Your Research Paper Overnight

A Markdown-only skill layer turns Claude Code and Codex into an overnight research team that drafts, critiques, and rewrites ML papers.

Repo, Full Access
  • ARIS is a Markdown-only skill layer for autonomous ML research, MIT-licensed at the repo
  • Uses cross-model adversarial review: Claude Code executes while GPT-5.5 xhigh critiques via Codex MCP
  • Ships 79+ composable skills covering ideas, experiments, review loops, paper writing, and rebuttals
  • Reported score jumps: 5.0 to 7.5 in overnight review, 4.0 to 8.5 across three writing rounds
  • Community submission to AAAI 2026 scored 7/10 from Stanford Agentic Reviewer as accept-worthy
  • Research Wiki gives persistent memory across sessions so failed ideas become anti-repetition memory

An open-source project called ARIS (Auto-Research-In-Sleep) has been quietly pulling in stars, and the pitch is bold enough to raise eyebrows: let two frontier LLMs argue with each other about your paper while you sleep, then wake up to a rewritten draft, a filled-in ablation table, and a reviewer score that has climbed a few points. The repo now sits at roughly 11.7k stars and 1.1k forks, and its authors claim an AAAI 2026 submission produced by the system pulled a 7/10 from Stanford's Agentic Reviewer with the note that it was a good paper worth accepting.

Two models, one adversarial loop

The design idea is deliberately narrow. ARIS chains 79 composable skills across the research lifecycle (literature and novelty, idea discovery, GPU experiments, autonomous review loop, paper writing, peer review) with cross-model adversarial review, where Claude executes and GPT-5.5 xhigh reviews via Codex MCP. The authors argue that using subagents from a single model to review its own output creates blind spots. Two is the minimum needed to break self-play blind spots, and 2-player games converge to Nash equilibrium far more efficiently than n-player ones. Adding more reviewers, they say, mostly adds API cost.

There is no framework to install. Every skill is a plain