Tag
This paper evaluates numerical representation invariance in language models, finding that evaluator interface issues can mimic reasoning failures and identifying model-specific errors like unit conversion problems in Mistral Small 4.