A benchmark finds agents can't yet redesign training algorithms

Researchers from Tsinghua University, Naver Labs and Einsia.AI published AI4AI-Bench, a benchmark testing whether AI agents can improve the training algorithms used to build AI systems — a capability at the centre of arguments about recursive self-improvement (RSI), the idea that an AI system able to improve the process that produces AI systems could accelerate its own development.

The benchmark comprises 10 frozen research repositories, each built around a distinct training-algorithm family. An agent is given four hours on a single Nvidia B300 GPU to read a repository’s code, modify the training algorithm, and test each idea against a proxy metric; whatever source-code patch it leaves behind is then applied in a fresh container and run from initialisation for up to twelve hours to produce a final score, normalised so that 0 represents the unmodified baseline and 1.0 represents a theoretical optimum. The authors said existing benchmarks in this space were typically won by collecting more data or tuning hyperparameters rather than by genuinely changing how a model learns, and designed AI4AI-Bench to isolate the harder capability specifically.

The best system tested reached a score of 0.25 — closing, in the authors’ words, “under a fifth of the distance” between the algorithm that was already there and the optimum — with most attempts failing to meaningfully alter how the underlying models learned at all. The result offers a data point against claims of imminent recursive self-improvement, arriving weeks after Anthropic’s essay “When AI builds itself” argued that the length of tasks models can complete unassisted has been doubling roughly every four months.