←Home

    Machines That Discover

    A guided tour of how AI systems are being built, trained and tested to make genuine discoveries, from open-ended search and idea generation to autonomous research agents and benchmarks built on unsolved problems.

    In order

    3 stories
    a scientist works at a stereo microscope with a fine probe, vintage colour
    a scientist works at a stereo microscope with a fine probe, vintage colour

    A scientist works at a stereo microscope with a fine probe, vintage colour · Photo: National Cancer Institute

    Discovery · Model-based search

    An LLM proposes designs and a statistical model picks which ones get tested

    In a 2026 preprint, Large Discovery Models pair LLM suggestions with a statistical model that weighs predicted value against uncertainty to decide what earns a costly test.

    arXiv preprint · 16 August 2026

    Photo: National Cancer Institute

    Discovery · Evolutionary search

    Training small language models on logged search steps from 355 optimisation tasks

    A 2026 preprint fine-tunes small open models on the score-improving steps of AI-driven evolutionary searches, with average gains on test problems that are uneven and unreplicated.

    arXiv preprint · 27 June 2026

    Photo: Fulvio Ciccolo

    Discovery · Evolutionary search

    Evolving web searches alongside solutions lifts AI optimisation for two of three models

    A 2026 preprint adds an evolving web-search loop to AI-driven evolutionary search, with large gains for one model, modest ones for another and losses for a third.

    arXiv preprint · 30 September 2026