ASCII art exposes LLM spatial blindness

The persistent inability of modern reasoning models to generate a simple "person touching eyes" proves that tokenization remains a spatial straitjacket that even 2026's best architectures haven't shed.

• –Tokenization is the primary culprit, as it translates 2D visual grids into 1D data streams, shattering the vertical alignment necessary for coherent ASCII structures.
• –The "ArtPrompt" vulnerability demonstrates that this spatial blindness is a security risk, allowing attackers to bypass safety filters using character-based obfuscation that models fail to "see."
• –While multi-modal models can describe images, their text-only inference engines lack a "mental canvas" to coordinate multi-line positioning, leading to the "hallucinated" chaos seen in recent benchmarks like ASCIIEval.
• –The success of models in 3D environments like MineBench suggests that while they can grasp abstract geometric logic, they lack the specific character-level spatial embeddings required for textual reconstruction.
• –Proprietary models like Gemini 3 and GPT-5 lead in recognition, but the generation gap remains high across the board, proving that "intelligence" and "spatial perception" are still not fully integrated in standard LLMs.