ARIS: Cross-Model Review and Persistent Memory Mechanisms for Reliable Long-Term Research Tasks

The MLNLP Academic Talk series invites researchers to share cutting‑edge techniques. This session features Yang Ruofeng, a PhD student at Shanghai Jiao Tong University, presenting ARIS (Auto‑Research‑in‑Sleep), an autonomous research system.

Yang Ruofeng, advised by Associate Professor Li Shuai, focuses on diffusion models (image, video, large‑scale language) and AI‑scientist agents. He is the primary author of the ARIS project (12 k stars), maintains a large community around autonomous research, has published eight papers at top conferences, interned at Tencent and Meituan, and received several prestigious awards. Personal homepage: https://wanshuiyin.github.io/.

Autonomous AI‑driven research systems aim for end‑to‑end paper generation, but single‑model self‑critique often produces relevance errors and gets trapped in local minima. ARIS addresses this by enforcing that the executor and reviewer belong to different model families, employing a three‑layer evidence audit chain and multi‑channel writing audit to systematically control the honesty of the research process. ARIS is fully open‑source and spans the complete pipeline from idea discovery to manuscript writing.

The presentation details the adversarial collaboration mechanism, the audit architecture, and the meta‑optimization outer loop, shares observations on the current limits of AI scientist capabilities, and outlines a vision for extending autonomous research to wet‑lab experiments.