A zero-dependency, Markdown-native autonomous ML research workflow system covering the full research lifecycle from idea discovery to rebuttal via cross-model adversarial collaboration.
ARIS decomposes the research process into 4 major workflows (Idea Discovery → Experiment Bridge → Auto Review Loop → Paper Writing) plus a standalone Rebuttal skill, each defined as pure Markdown readable by any LLM. The core mechanism employs cross-model adversarial collaboration—Claude Code executes experiments and writing while external models like GPT-5.4 independently review, receiving only file paths to prevent information contamination. A four-layer Evidence & Claim assurance stack (experiment-audit → result-to-claim → paper-claim-audit → citation-audit) and Assurance Gate enforce mandatory audits under --effort: beast mode with non-zero exit codes blocking final reports.
Research Pipeline#
| Workflow | Command | Input → Output |
|---|---|---|
| W1 — Idea Discovery | /idea-discovery | Research direction → Idea report + Experiment plan |
| W1.5 — Experiment Bridge | /experiment-bridge | Experiment plan → Runnable code + Experiment log |
| W2 — Auto Review Loop | /auto-review-loop | Paper + results → Iteratively improved paper |
| W3 — Paper Writing | /paper-writing | Narrative report → Structured LaTeX paper + PDF |
| W4 — Rebuttal | /rebuttal | Paper + reviews → Character-limited response |
| Full Pipeline | /research-pipeline | Research direction → W1→W1.5→W2→W3 chained output |
Cross-Model Adversarial Collaboration#
- Executor: Claude Code / Codex CLI — writes code, runs experiments, drafts papers
- Reviewer: GPT-5.4 / Gemini / GLM — critiques, scores, requests revisions
- Core Constraint: Executor and reviewer must be from different model families; reviewers receive only file paths, never executor summaries
Quality Assurance#
- Four-layer assurance stack: experiment-audit → result-to-claim → paper-claim-audit → citation-audit
- Assurance Gate:
--effort: beast+--assurance: submissionenforces submission-stage audits - Rebuttal safety gates: no fabrication / no over-promising / full coverage
Paper Output Tools#
/paper-slides (Beamer PPT), /paper-poster (poster PDF/PPTX/SVG), /paper-illustration (figure generation)
Persistence & Collaboration#
- Research Wiki: optional persistent project memory shared across skills
- Overleaf bidirectional sync via Overleaf Git Bridge
Compatible Environments#
Claude Code, Cursor, Trae (ByteDance), Antigravity (Google), Windsurf, Codex CLI, OpenClaw, etc. GPU backends: local / remote / vast / modal.
Zero-Dependency Design#
The entire system consists of pure Markdown files—no frameworks, databases, Docker, or daemons. Skills communicate via chained plain-text artifacts: IDEA_REPORT.md → EXPERIMENT_PLAN.md → EXPERIMENT_LOG.md → NARRATIVE_REPORT.md → paper/main.tex → paper/main.pdf.
Validated Output#
Multiple papers assisted by ARIS have been produced, including an accepted AAAI 2026 Main Technical Track paper.
Unconfirmed Information#
- Whether ARIS itself has an accompanying technical report or paper (not explicitly stated)
- Author wanshuiyin's institutional or academic background (not annotated)
- Detailed configuration steps for ModelScope free-tier usage (not specified)
- Independent website and HuggingFace page (not found)