An SI agent with a "frozen" model scored 100 out of 100 on all 25 public ARC-AGI-3 games by rewriting its own rulebook
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An agent may improve not by retraining the model but by rewriting its own rules and memory, though for now the result is claimed only by the authors.
Authors from University College London and Huawei's Noah's Ark lab (Memento 3 preprint, arXiv 2610.11794) claim their agent cleared all 25 public ARC-AGI-3 games with a mean relative action-efficiency score of 100.0, using 44% of the human action count (7,518 versus 17,135), without fine-tuning the language model. The agent keeps a natural-language "rulebook" of testable hypotheses about the environment and compiles it into a code module that it rewrites after each wrong prediction. According to AI Weekly, a baseline agent on Claude Opus 5 scores 40.7 on the same measure. The result was obtained on the public games and is claimed by the authors.
Sources 2