Proc‑gen meets LLMs

Procedural generation is the old trick of letting code build game worlds; large language models are now being bolted on to name places, write lore, and draft quests on top of those maps. (arxiv.org)(docs.unity3d.com) The terrain part usually starts with noise, a math pattern that works like controlled static and turns flat grids into hills, valleys, and coastlines. Unity’s Terrain Tools package includes a Noise Editor for exactly that workflow, and many hobby projects still rely on Perlin noise as the base layer. (docs.unity3d.com)(github.com) The new wrinkle is that developers are treating the language model as a planner above that terrain stack instead of as the terrain generator itself. A recent GitHub project called Confluent AI describes a C++ and OpenGL system that keeps Perlin-noise terrain underneath while using a large language model to interpret natural-language changes in real time. (github.com 1)(github.com 2) That split matches where research has been moving over the past year. An October 2024 survey on procedural content generation said large language models were changing the field, and later papers pushed them into editable 3D assets and story-to-scene pipelines instead of asking them to brute-force entire worlds from scrat(arxiv.org 1)(arxiv.org 2)(arxiv.org 3)81)) One January 2026 paper, Proc3D, described a graph-based system for editable 3D models that accepts natural-language edits and reported more than 400 times faster editing than full regeneration in its tests. Another 2025 paper turned short narratives into tile-based game scenes by extracting object relationships from model-writ(arxiv.org 1)(arxiv.org 2)09.04481)) Quest generation is following the same pattern. Public code repositories now describe pipelines that pair noise-based world maps with language-model systems that fill those spaces with location descriptions, role-playing game quests, and other narrative hooks tied to the generat(github.com 1)(github.com 2)t-generation)) The appeal is practical: the deterministic parts still handle shape, scale, and repeatability, while the language model handles text-heavy tasks that used to require hand-written content. That lets a developer keep a reusable world seed for terrain while swapping in different quest lines, factions, or place histories on top. (github.com)(arxiv.org) The tradeoff is reliability. Academic and hobby projects alike still describe open problems around coherence, latency, and control, especially when generated story text has to line up with playable spaces and rules the game can actually enforce. (link.springer.com)(arxiv.org) That is why the most convincing demos right now tend to be hybrids. The code still lays down the mountain, corridor, or puzzle room; the model is increasingly being asked to explain it, populate it, or give the player a reason to care what is on the other side. (arxiv.org)(github.com)