Assessing Consciousness-Related Behaviors Using the Maze Test

Rui A. Pimenta, Tim Schlippe, and Kristina Schaaff evaluate twelve models on text-described mazes requiring first-person navigation instructions. Thirty-four test mazes follow separately allocated demonstrations.^[1]

Method and contribution

Complete-path accuracy scores fully correct solutions; partial-path accuracy counts consecutive correct steps before the first error. Performance varies by model and prompting condition. These measures distinguish starting a valid plan from carrying it through.^[1]

Alignment relevance and limits

The results inform reliable sequential behavior and Machine Introspection discussions. The authors associate these capabilities with components of consciousness, but their ethical statement explicitly says the test does not measure consciousness itself. Neither success nor failure proves the presence or absence of phenomenal experience. Claims about persistent self-models remain an interpretation of behavioral scores. This paper differs from MazeEval, which uses interactive coordinate feedback rather than a complete text-described maze.