Tag
This paper introduces a round-trip protocol to assess how well tree-structured expressions survive serialization into natural language by language models, revealing that the process is lossy, asymmetric, and trainable with fine-tuning.