What makes a world model? A four-part test that foundation models, digital twins fail
marinkazitnik · x · 2026-09-18
Marinka Zitnik proposes one test for world models: does the model represent a system state, accept a specified action, predict the resulting state, and continue simulating from that prediction? By this standard, foundation models can serve as encoders but aren't world models; digital twins mirroring only current state don't qualify either. Virtual cells, embryos, and patients remain aspirational, each exposing a gap: virtual cells often fail to beat linear baselines in poorly covered contexts; embryo atlases observe different cells per stage with few interventions; xHealth records span years but treatments are selected, not randomized.
Related event: Harvard team proposes biomedical world models framework in Cell(8 posts)→
More from Research
- ML is leaving its alchemy era: assumed truths can now be instrumented and falsified — mike64_t · 2026-09-18
- Together AI paper: score centering cancels drift in off-policy RL under train-inference mismatch — PandaAshwinee · 2026-09-18
- Cell's AI-in-biology special issue debuts LongevityBench, a 17-task aging benchmark — JosephJacks_ · 2026-09-18
- After OpenAI cracks Navier-Stokes, the era of the human mathematician is ending — yeastsplainer · 2026-09-18
- Custom BB-SLAM hits 0.02m accuracy with pure visual odometry, no LiDAR or IMU — broodsugar · 2026-09-18
- AI4Science Breakthrough Said to Exceed Expectations, Could Transform Computational Chemistry — BenBlaiszik · 2026-09-18