arxivcs.AI2026-07-01
AGI Maze as a Benchmark Framework for World-Modeling Agents
Large language models (LLMs) are powerful pattern-completion systems, but their default operating mode - predicting the next token from a static context - does not reliably produce persistent, manipulable representations of an external world. Many tasks that look like "reasoning"…