JEPA-TTT: Persistent test-time training lifts world model planning by 153%
DanielKhashabi · x · 2026-10-03
Honda Research Institute and Johns Hopkins University released JEPA-TTT (World Models in Physical AI Workshop @ NeurIPS 2026).
- Problem: a pretrained latent world model's predictions become unreliable when test-time dynamics differ from training (dynamics shift).
- Method: JEPA-TTT persistently applies self-supervised updates to the latent dynamics predictor of a pretrained action-conditioned JEPA world model, accumulating adaptation across episodes while the visual encoder and reward head stay frozen, preserving pretrained representations and task objectives. Planning needs neither goal images nor online environment rewards. It uses dense replay: prediction windows at every temporal offset are stored in a growing buffer and sampled for minibatch updates.
- Results: improves planning on all 8 dynamics shifts across 4 continuous-control environments; after 500 test-time episodes, autoregressive latent prediction error drops 83% on average and planning performance improves 153% over the frozen JEPA world model.
More from Research
- Researchers call for simplest-possible baselines in ML×Bio papers over flashy ablations — anshulkundaje · 2026-10-03
- Meta's Muse Spark helps disprove an evolution algebras conjecture, one of six AI-math papers — AIatMeta · 2026-10-03
- Meta partners with mathematicians on six papers, solving five open research problems with Muse Spark — AIatMeta · 2026-10-03
- PDFs often lie about or omit word and character positions — 5 failure modes explained — VikParuchuri · 2026-10-03
- NBER paper: technical change 'commoditizes' labor, boosting productivity but suppressing wages — PeterHndrsn · 2026-10-03
- bycloud breaks down Jev: how can a model get better without generating words? — bycloud · 2026-10-03