Debate over world-model gradients: zero-order wins on speed

Researchers debate whether first-order gradients through learned world models are worth it, with one arguing the payoff is limited when samples are cheap; another demos a footed locomotion policy trained in about 30 seconds using zero-order policy gradients.

2026-10-26 ~ 2026-10-26 · 2 related posts

• First-order world-model gradients give minor training speedups over zero-order PG — breadli428 · 2026-10-26
• Locomotion policy trained in under 30 seconds with zero-order policy gradient — breadli428 · 2026-10-26