中文 Apply now

News

Releases, programs and progress

The running record: DFM and module releases, the Scientist Collaboration Program, project progress and team news.

Release

DiscoveryRSI · recursively learning to discover valuable hypotheses: research agents that improve from their own research experience

DiscoveryRSI lets research agents improve recursively from their own research experience: every revision of a hypothesis is kept with the evidence and decisions behind it, and that record trains the model, refines the workflow and opens new investigations. Against research published after the model cutoff, 6.3% of DiscoveryRSI-27B's hypotheses have a concrete connection a researcher could build on, the highest among the models compared.

Read more →

  1. Release

    RSI-Anything · recursive self-improvement across domains: world models in every round of improvement

    RSI-Anything puts world models inside the recursive self-improvement loop: predictions pick which candidates are worth testing for real, observed feedback improves the policy or workflow, the world model is updated too once its predictions prove unreliable, and only updates that pass their checks carry into the next round. In code repair with the coder held fixed, success rises from 26.3% with a static verifier to 29.2% with predictive selection and 31.2% after world-model adaptation.

    Read more →

  2. Release

    JEPA-Anything · cross-domain unified science world model: orthogonal predictive factorisation for latent world models

    JEPA-Anything splits latent targets into complementary factors with orthogonal predictive factorisation (OPF), and is evaluated across vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. Against matched JEPA baselines, it improves reported metrics on all 10 dynamics tasks.

    Read more →

  3. Release

    ScienceIDE · science integrated development platform: turning real scientific code into verifiable environments

    ScienceIDE turns real scientific code, research procedures, task goals and verification criteria into scientific environments that run, measure and reproduce. The PhAI-IDE model family, trained on verified interaction trajectories from these environments, shows gains in held-out scientific-code repair and across selected general-purpose benchmarks. Current task forms are accelerate, discover, repair, reproduce and integrate, and task proposals are open.

    Read more →

  4. Release

    ScienceBuddy · AI research partner: an interactive Scientific Agent and recursive self-improvement in real research

    ScienceBuddy is an independent research and engineering exploration built around interaction with scientists, research feedback and the continual improvement of agents: follow-ups, edits, acceptances, rejections and re-runs become feedback that keeps the research agent's workspace evolving.

    Read more →

  5. Release

    From solving research tasks to taking part in discovery: PhAI Labs proposes the DFM paradigm

    PhAI Labs proposes DFM, a research paradigm and systematic account of Discovery Intelligence, and opens the DFM Scientist Collaboration Program. Over the following three days ScienceBuddy, ScienceIDE and JEPA-Anything, three independent works, are released in turn.

    Read more →

  6. Program

    Bring real scientific questions into next-generation AI: the DFM Scientist Collaboration Program opens

    The DFM Scientist Collaboration Program brings real scientific questions, high-value data, interactive research environments and real verification conditions into DFM. The first cohort prioritizes life science and biomedicine, while remaining open to chemistry, materials and other frontier fields.

    Read more →