MLS-Bench-Lite: Results as Published

MLS-Bench asks whether an AI system can invent generalizable ML methods rather than only tune existing ones. The full benchmark has 140 tasks across 12 domains, and each task asks an agent to propose one modular improvement to a targeted component, such as a loss, an attention variant, a sampler or a routing rule, then shows the change holds across models, datasets and seeds. MLS-Bench-Lite is the official 30-task subset shown on the leaderboard, and results are compared with a human reference labelled Human SOTA.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.