Geração de mapas 2D com LLMS e validação baseada em agentes

Abstract

This paper investigates LLMsforprocedural generation of 2D dungeoncrawler maps, combiningprompt-guided creativity with deterministic agent-based validation. We propose a hybrid pipeline: (i) strict JSON-based ASCII grid generation with explicit dimension/entity/topology rules; (ii) grid normalization and wall auto-tiling; and (iii) screening via an A*-based validator for syntactic, connectivity, and accessibility constraints. A Godot 4.6 prototype was evaluated across campaigns with varying map sizes and entity densities, comparing Gemma 4, Ministral 3, and Gemini 3.1 Flash Lite. Raw LLM outputs proved unstable, with critical failures driven by spatial blindness and poor counting; however, the retry-and-validate loop recovers playable layouts and mitigates soft-locks. A Likert study suggests good thematic coherence (notably for local models), while inference latency limits real-time use, favoring offline pre-production.