LeCun's lab borrows a brain trick to double AI goal-reaching success to 94%
alex_verem · x · 2026-09-26
- A paper from NYU, Brown and University of Toronto (ICML 2026), with Yann LeCun as co-author, improves latent planning in JEPA world models using "temporal straightening," inspired by perceptual straightening in human vision.
- The problem: world-model latent maps are warped—two points that look close can be far apart in reality, so the AI walks into walls.
- The fix: a curvature regularizer penalizes bending in latent trajectories, making Euclidean distance a better proxy for geodesic distance.
- Results: two-room goal-reaching jumped from 52.7% to 90.7%; U-shaped maze from 44% to 94%; both hit 100% with re-planning. A simple planner on the straightened map matched a much heavier planner and ran 10x faster.
- Limitations: tests remain small-scale (mazes, block pushing), and errors compound over long plans. It's a concrete step for LeCun's world-models-over-chatbots thesis.
More from Research
- Scandinavian Journal of Statistics editors reflect on publishing in the age of generative AI — michaelchchoi · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- Running Qwen2.5-Coder-3B on Android: Jupyter + Pydantic Evals on-device eval pipeline — alex_b3b · 2026-11-11
- cloneofsimo releases a video explaining determinantal point processes — cloneofsimo · 2026-11-11
- Mapping Qwen3-8B's internal touch body map into an anatomical plate — repligate · 2026-11-11