Quantifying Information Loss in LLM Chain-of-Thought Serialization
September 28, 2026
A study investigates how much tree-structured compositional content is lost when LLMs serialize reasoning into natural language. Using a round-trip protocol of arithmetic expression generation and extraction, the research measures the communication bottleneck inherent in text-based reasoning.
HOW THIS AFFECTS YOU
●
builderConsider using structured formats like JSON instead of free-text for complex reasoning tasks to avoid this bottleneck.
●
researcherThis provides a framework for measuring structural information loss in language models.