The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models

Apple researchers have introduced a round-trip protocol to measure how much tree-structured information is lost when language models serialize complex data into natural language. This study investigates the communication bottleneck inherent in chain-of-thought reasoning, where models must convert structured logic into standard text. By quantifying this data degradation, the findings provide a framework for improving the accuracy of models that rely on compositional reasoning and intermediate logical steps.