ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration
Open-source ARIS harnesses cross-model critics to catch unsupported claims in AI research.
Long-running AI research agents often fail not by crashing, but by producing plausible results built on incomplete evidence. ARIS (Auto-Research-in-sleep) addresses this through adversarial multi-agent collaboration: an executor model drives research forward, while a reviewer from a different model family critiques intermediate artifacts and requests revisions. This cross-model setup acts as a built-in fact-checker, ensuring claims are backed by solid evidence before they pass through the pipeline.
The system's three-layer architecture includes an execution layer with 65+ reusable Markdown-defined skills and a persistent research wiki, an orchestration layer coordinating five end-to-end workflows with adjustable effort settings, and an assurance layer with three-stage verification (integrity check, result-to-claim mapping, claim auditing) plus a five-pass scientific editing pipeline and mathematical-proof checks. ARIS also features a prototype self-improvement loop that records research traces and proposes harness improvements—but only after reviewer approval. Open-source and available on GitHub, ARIS is a step toward autonomous research that doesn't just generate outputs, but verifies them rigorously.
- Uses adversarial collaboration: an executor model (e.g., GPT-5) and a reviewer from a different model family (e.g., Claude) to catch unsupported claims.
- Includes 65+ reusable skills, five end-to-end workflows with adjustable effort, and a 3-stage assurance process for evidence verification.
- Features a self-improvement loop that records research traces and suggests harness upgrades, approved only by the reviewer model.
Why It Matters
Automates rigorous AI research validation, saving researchers from wasted effort on false or unsupported findings.