Skip to content
Trending storyPast story

600 experiments probe LeWM's long-horizon planning limits and fixes

1 report1 sourceupdated 3 days ago

What happened

AI digest

Haddy Alchaer ran 600 experiments on the JEPA world model LeWM and found four planning methods that need no extra training, raising success on long-horizon tasks from 22% to 85%, Hacker News reported on November 14, 2026. The methods set subgoals every 25 steps, add a cost for jumps across demonstrations in the memory graph, fill memory with transformed demonstrations, and adjust goal scoring. On a harder test where start and goal come from different demonstrations, the best planner reached 56%, below the long-horizon result.

Written by AI from the coverage · updated 3 days ago

Coverage

Follow the reports to see the story from different sides.

Nov 14
  1. Hacker News front page
    After 600 experiments, here's what I discovered about JEPA world models

    Haddy Alchaer 对 LeWM 进行了 600 次实验,找到四种无需额外训练的规划方法,将长时程任务成功率从 22% 提升至 85%。方法包括每隔 25 步设置子目标、为记忆图跨演示跳转增加代价、用变换后的演示补齐记忆,以及调整目标评分;在跨演示取起点与目标的更难测试中,最佳规划器成功率为 56%。

Heat over time

Heat now 3·Comparable peak 9(Nov 14 05:00)·Comparable change over 24 hours -49%

02.557.510Nov 1405:00Nov 1420:00Nov 1510:00Nov 1601:00

The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.