Platonic Representation Theorem questioned: positive-pair choice embeds worldview assumptions
AnnaCiaunica · x · 2026-11-03
A new result (with Yann LeCun) upgrades the Platonic Representation Hypothesis to a theorem: JEPAs trained separately on any modality provably learn the same representation up to rotation, with a 109M JEPA-Text aligning to vision better than 72B LLMs. Yanbo Zhang pushes back: the theorem assumes both training runs share the same finite set of positive examples, whose selection reflects the researcher's worldview. He also invokes Kleene's recursion theorem and the undecidability of functional equivalence as a potential fundamental obstacle.
More from Research
- WorldFlow: generative world model jointly predicting video and 3D trajectories, ADE down 54.2% — Yuhao Lu · 2026-11-03
- FLAMIA: MXFP8 microscaling attention kernel for fast, causally-safe training on Blackwell — yresearch · 2026-11-03
- Lab2Sim benchmark: turning 218 real experiment videos into interactive virtual experiments — PolyUHK · 2026-11-03
- Tsinghua study: continual harness self-evolution causes severe catastrophic forgetting in coding agents — Shangziqi Zhao · 2026-11-03
- PRIMA animal 3D mesh recovery accepted to WACV2027 with 9K pseudo-3D dataset — TrackingActions · 2026-11-03
- AI for Science event spotlight: physics-informed losses and quantum chemistry — burny_tech · 2026-11-03