LLM Zelda Level Generator Hits Diversity Metric but Produces Visually Boring Rooms

Amidos2006 · x · 2026-09-09

The author shares a paper on LLM-driven game level generation (PCG), co-authored with uTheProdigy, MatthewSiper, Steven James, and Togelius.

An honest limitation: high fitness ≠ visual diversity. Their Zelda generator learned to clear the room and shuffle key/door/player positions — scoring well on shortest-path diversity metrics while being boring to look at.

The proposed next step is quality-diversity (QD) search, to maintain genuinely diverse outputs while still optimizing objectives.

Related event: LLM-Generated Zelda Levels Hit Metrics but Look Boring(2 posts)→

Original post →

More from Research

Research channel →