ASI‑Evolve automates research

A new paper on arXiv makes a blunt claim: AI can now automate meaningful chunks of AI research itself. The system is called ASI-Evolve, and its creators say it can generate ideas, run experiments, analyze the results, and feed those lessons back into the next round. The paper appeared on March 31, 2026, from researchers at Shanghai Jiao Tong University, Shanghai AI Laboratory, and GAIR. (arxiv.org) That matters because most AI progress still comes from a slow loop. Researchers propose an architecture, or a data-cleaning rule, or a training method. Then they implement it, test it, inspect the failures, and try again. ASI-Evolve is built to close that loop with less human intervention. Its core design is simple enough to describe in one sentence: learn, design, experiment, analyze, repeat. But the paper’s real point is that the loop does not just run automatically. It also stores what it learns. (arxiv.org) The system has two pieces that make it more than a brute-force search engine. One is a “cognition base,” which acts like a memory bank of human priors and accumulated lessons. The other is an analyzer that reads experimental outcomes and turns them into reusable guidance for future rounds. That is the self-improving part. Instead of treating each failed run as wasted compute, ASI-Evolve tries to convert failure into strategy. (arxiv.org) The strongest result came from neural architecture search. The team says ASI-Evolve discovered 105 state-of-the-art linear attention architectures. Its best model beat DeltaNet by 0.97 points, which the authors describe as nearly three times the gain delivered by recent human-designed improvements in the same area. That 3× figure is the headline number floating around online, and in the paper it refers to the size of the improvement over a prior baseline, not to a universal claim that the system is three times better than human researchers at research in general. (arxiv.org) That distinction is important because the paper is ambitious, but it is not magic. The authors did not show a machine replacing an entire research lab. They showed a framework performing well on three bounded but central parts of the AI stack: model architectures, pretraining data curation, and reinforcement-learning algorithm design. Inside those lanes, the results are still striking. In data curation, the evolved pipeline improved average benchmark performance by 3.96 points, with gains above 18 points on MMLU. (arxiv.org) Then the paper turns to reinforcement learning, where the system was asked to invent better algorithms rather than just tune existing ones. There, the discovered methods beat GRPO by up to 12.5 points on AMC32, 11.67 on AIME24, and 5.04 on OlympiadBench. Those are not tiny nudges from hyperparameter cleanup. They are large enough to suggest the system is finding non-obvious changes that survive contact with evaluation. (arxiv.org) The broader story is not that AI suddenly became a scientist in the romantic sense. It is that one of the most expensive parts of modern AI development may be turning into a search-and-memory problem that machines can increasingly handle on their own. The authors even say they see early signs that the same pattern could transfer beyond AI, with initial experiments in mathematics and biomedicine. (arxiv.org) The concrete thing to watch is not the paper title. It is the repository. The team says ASI-Evolve is fully open-sourced, and a related project page for its architecture-discovery system says it released more than 100 discovered linear-attention designs, along with the pipeline, database, and cognition library used to generate them. (arxiv.org)