• Long-horizon planning through a world model suffers exploding/vanishing gradients because the Jacobian with respect to early actions is a product of per-step Jacobians whose conditioning scales exponentially with the planning horizon.
• GRASP's key trick is stopping gradients into the state input of the world model, trusting only action gradients, because states sit on a low-dimensional manifold with adversarial-example-like brittleness in directions orthogonal to it, which the action space doesn't share.
• Because the naive stop-gradient objective has a trivial minimum (stay put, then jump to the goal on the last step), GRASP adds a dense per-timestep goal term to every predicted state, balanced against the stop-gradient dynamics loss.
• On the Push-T benchmark at horizon 80, GRASP hit a 10.4% success rate in a median 58.9s, versus CEM's 2.8%/132.2s and LatCo's 0.0%; at horizon 40, GRASP matched CEM's roughly 60% success rate in about a quarter of the time (8.5s vs 35.3s).
• GRASP periodically breaks from its fast lifted-state optimization to run a short refinement phase using the true serial rollout gradient, to keep the lifted trajectory grounded in dynamics actually reachable by a real rollout.

• benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
• grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.

It solves a specific practical bottleneck (brittle gradients through high-dimensional vision models) that has limited long-horizon planning with learned world models.

• articleARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World ModelsZhengshu Zhang, Zhiyuan Li
• articleGSPO: Towards Scalable Reinforcement Learning for Language ModelsQwen Team
• videoScaling to Long Horizons — Ross Taylor & Chengxi Taylor, General ReasoningAI Engineer

Checking sign-in…

Loading comments…