Best For
ML researchers who want to compress the idea-to-submission timeline. Especially useful for PhD students, postdocs, and independent researchers who lack frequent peer feedback. If you have a research question and GPU access, ARIS can run literature review, experiment design, data analysis, manuscript writing, adversarial review, and revision cycles overnight — you wake up to scored drafts with identified weaknesses.
How I Actually Use It
Set up a research question, seed literature, and experiment parameters. ARIS assigns Claude Code as the executor (reads files, writes code, runs experiments, compiles results) and GPT as the adversarial reviewer (scores the manuscript, identifies logical gaps, suggests improvements). The two models never self-review — Claude's work is always reviewed by GPT, and vice versa.
The system runs eight core skills covering the full lifecycle: literature review, experiment execution (with GPU bridge for remote compute), manuscript writing, cross-model review, revision, citation verification, claim auditing, and submission preparation. Skills communicate through Markdown artifacts — no database, no framework, no Docker.
Where It Is Strong
- Cross-model adversarial review is theoretically grounded (Adversarial Bandit framework), not just "ask another model to check." The reviewer never sees the executor's reasoning, only the output and its contract
- Three-layer academic integrity: citation audit (verify every reference exists), claim audit (check every claim has evidence), experiment integrity (ensure reproducibility). These catch real problems, not just style issues
- Nightmare Mode: GPT reads the entire repository directly and stress-tests the manuscript as a hostile reviewer. The closest simulation to a Reviewer #2 you'll get from AI
- Zero dependency principle. Pure Markdown skills plus Python utility scripts. Swap models freely — supports Claude, GPT, Gemini, DeepSeek, MiniMax, and even free ModelScope endpoints
- Context survival: REVIEW_STATE.json persists workflow state across context compaction events, so long overnight runs don't lose progress
Where It Fails
- Dual subscription cost. Full functionality requires Claude Code Pro and ChatGPT Plus/Pro. Budget roughly $40-60/month minimum
- GPU required for experiments. The Experiment Bridge needs SSH access to remote GPU (Vast.ai, Modal, or your own servers). Without GPU, you're limited to literature review and writing
- AI review scores skew optimistic. Because ARIS iterates against AI reviewers, AI scores trend upward. Real human reviewers will likely be harsher. Don't treat an 8/10 AI score as guaranteed acceptance
- High barrier to entry. You need to understand ML research workflows to use this effectively. Not a tool for casual users
- Context window pressure. Extended auto-review loops in beast mode can trigger compaction, though the state persistence mechanism mitigates this
Pricing, Difficulty, and Risk
The framework itself is free (MIT license). Real cost is the LLM subscriptions: Claude Code Pro ($20/mo) plus ChatGPT Plus/Pro ($20-200/mo). GPU costs vary. Difficulty is high — you need ML domain knowledge, comfort with remote compute, and understanding of the research publication process. Privacy risk: your research data goes through both Claude and GPT APIs. Integrity risk: AI-generated research must be disclosed per venue policies.
Verdict
The most complete open-source framework for autonomous ML research. The cross-model adversarial review is a genuine advance over single-model self-assessment. Worth the dual subscription cost if you're actively publishing. But calibrate expectations: ARIS accelerates research, it doesn't replace the researcher. Human judgment on research direction, significance, and ethical implications remains irreplaceable.
Source
- GitHub: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep
- Technical Report: arXiv:2605.03042