JEPA-like world models collapse on distractors and natural video, researchers report
inductionheads · x · 2026-09-27
@bonniesjli responded to a question about whether JEPA has ever been validated on standard RL tasks: the team compared against a Momentum Prediction baseline (similar to JEPA) and found it comparable in clean settings but collapsed under distractor and natural video conditions. The takeaway: JEPA-like world models need more work to function well on these tasks. The original questioner asked whether JEPA has ever been shown to work on DM Control or benchmarked head-to-head against standard model-based RL — this reply suggests such evidence remains scarce.
More from Research
- kalomaze proposes testing which nanogpt tricks survive causal NTP over DCT coefficients — kalomaze · 2026-09-27
- AI-designed drug turns back 6 aging clocks; semaglutide extends mouse lifespan 12% — rand_longevity · 2026-09-27
- Could perfect monosemanticity enable training coding agents without verifiers or RL? — menhguin · 2026-09-27
- Local agent eval: challenger model wrote tool calls as prose 5/80 times — a silent failure class — Grimmoner · 2026-09-27
- TalkPlayData-backed conversational music recsys challenge at RecSys 2026 draws 41 teams — keunwoochoi · 2026-09-27
- A better metaphor for LLMs: bags of contextually activated circuits and heuristics — xuanalogue · 2026-09-27