AI

JEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning under Dynamics Shifts

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 00:18:24 UTC.

Briefings

JEPA-TTT adapts a pretrained world model by updating its latent dynamics predictor across test-time episodes. The method uses dense replay to form prediction windows at every temporal offset, accumulating self-supervised updates while keeping the visual encoder and reward head fixed. This approach improves planning performance by 153% over a frozen baseline after 500 test-time episodes, reducing prediction error by 83% on average across eight dynamics shifts in four continuous-control environments.

Key points

• JEPA-TTT adapts a pretrained world model by updating its latent dynamics predictor across test-time episodes using dense replay.
• The method improves planning performance by 153% over a frozen baseline after 500 test-time episodes.

Context

World models enable agents to plan by predicting future states, but their predictions can become unreliable when test-time dynamics differ from training. JEPA-TTT addresses this by adapting the latent dynamics predictor throughout test time using dense replay, which forms prediction windows at every temporal offset.

Limitations

Results are based on eight dynamics shifts across four continuous-control environments. The study does not report results for discrete-control environments or other specific dynamics types. The 500 test-time episode limit may limit generalizability to longer planning horizons.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/10/2026, 05:28:41 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeickvsswff4k63yftzwysnlieuwg4evo3hqdvtiu24w2mpx2zuqciu