Parsed, Plausible, Wrong: Output Format Controls Silent Errors in Small-Model Semantic Parsing

In 30 Sekunden

Was passiert ist

Small language models produce grammatically valid but semantically wrong parse trees silently. Using natural language output format instead of symbolic notation cuts silent errors by 2.3x.

Warum es zählt

You deploy fine-tuned small parsers for structured tasks like order processing where wrong outputs go undetected downstream without explicit validation.

Achtung

The study uses one benchmark (pizza ordering). Results may not generalize to other semantic parsing domains or different model architectures and sizes.