Autonomous architecture search spotted
A public thread highlighted ASI‑Evolve, an AI that autonomously searches for new model architectures but still requires human oversight to avoid unsafe or fragile designs (x.com). Analysts warned that these self‑improving systems introduce new risks because they can explore unexpected model topologies and behaviors that humans must monitor (x.com).
AI researchers have posted a new system, ASI-Evolve, that lets an agent search for model designs on its own instead of testing only human-written ideas. (arxiv.org) The March 31, 2026 paper says ASI-Evolve runs a repeated loop — learn, design, experiment, analyze — and uses stored prior knowledge plus an analyzer module to guide later rounds. The authors say that setup let it work across data curation, model architecture, and learning algorithm design. (arxiv.org) In the architecture part, the paper says the system generated 1,350 candidates across 1,773 experiments and found 105 state-of-the-art linear-attention architectures. Its best reported model beat DeltaNet by 0.97 points, according to the arXiv abstract and HTML version. (arxiv.org, arxiv.org) Model architecture search is the process of trying different internal layouts for a neural network — the way an engineer might test many bridge blueprints before pouring concrete. Earlier neural architecture search systems usually explored spaces that humans defined in advance. (arxiv.org) The newer papers around ASI-Arch and ASI-Evolve describe a broader setup: an agent proposes a design, writes or modifies code, runs training jobs, reads the results, and proposes the next design. The ASI-Evolve repository says three roles drive that loop: Researcher, Engineer, and Analyzer. (arxiv.org, github.com) That shift has drawn attention because the system is not only tuning numbers inside a fixed recipe; it is also changing the recipe itself. The 2025 ASI-Arch paper said older neural architecture search was limited to “human-defined spaces,” while its system aimed at “architectural innovation” beyond those boundaries. (arxiv.org) The same research line has also drawn skepticism from reviewers. OpenReview records show an ICLR 2026 submission on the architecture-discovery framework was rejected after reviewers said gains were modest or inconsistent, ablations were missing, and comparisons with other automated methods were not rigorous enough. (openreview.net) Risk researchers have been warning that more autonomous agents create failure modes that do not show up in a one-shot chatbot. A 2025 survey from Tsinghua University researchers lists memory poisoning, tool misuse, reward hacking, emergent misalignment, and “irreversible tool chains” among the risks tied to autonomy, memory, and recursive planning. (arxiv.org) That leaves a narrower claim than the hype around “self-improving AI.” The papers and code show a system that can automate long research loops and search strange new topologies at scale, while the review record and security literature show why humans still have to check whether those designs are robust, reproducible, and safe. (arxiv.org, openreview.net, arxiv.org)
