DeepSeek R1 Distill falls short on grids

This is less a “bad model” story than a reminder that structured spatial reasoning is a different beast from generic chat or math reasoning, and 7B distills often hit that wall fast.

• –A 10x10 board plus 50 legal actions is a constrained planning problem; the model has to preserve exact state, not free-associate
• –The hardware ceiling points toward a larger quantized model, not another 7B-class distill, if the goal is fewer hallucinations
• –Recent grid-world research suggests spatial performance is highly representation-dependent, so raw board formatting can matter as much as model size
• –For a text-adventure engine, state encoding and action masking may buy more reliability than a model swap alone