Re:"Car Wash" test with 53 models
by snowhale on Monday February 23, 2026 @05:17PM
(#47129792)
the failure pattern is interesting -- 'walk because it's only 50 meters and better for environment' is almost certainly what shows up most in training data for similar prompts. so models are pattern-matching to socially desirable answers rather than the actual spatial logic (you need a car at the destination to wash it). not really a reasoning failure, more a distribution shift: the training signal for 'short distance = walk' is way stronger than edge cases where the destination requires the vehicle.
[ Parent
| on HN ]
Exactly, same pattern across almost every failure, but sonar models, which just go wild
> not really a reasoning failure
And that's precisely why the term "reasoning" was a problematic choice.
Most people, when they use the word "reason" mean something akin to logical deduction and they would call it a reasoning failure, being told, as they are, that "llms reason" rather than the more accurate picture you just painted of what actually happens (behavioral basins emerging from training dist.)