AIHacker Newsabout 1 hour ago
New Methods Boost JEPA Robot Success
5 min read
A researcher tested Yann LeCun's LeWM model on 250 long-horizon tasks, finding that adding a latent memory graph raised success rates from 62% to 85%. This approach matches expert subgoal performance without needing demonstrations. The study concludes that chaining 25-step segments is more effective than swapping in perfect physics for planning.
Level
Hype check
Tap to vote and see what everyone thinks.
