Sign in
← AI PulseSep 22, 2026

Deep · research · Single-source brief

Tree-Structured Expression Serialization in Language Models Shows Lossy, Asymmetric Communication

A round-trip study on arXiv cs.AI reveals that serializing tree-structured expressions into natural language and then extracting them is a lossy and asymmetric communication channel for language models, with failures primarily originating during generation.

By Illumora Editorial

Source · Sep 22, 2026, 4:00 AM · On Illumora · Sep 22, 2026, 4:03 AM

Media from the primary source — shown here so you can stay on Illumora.

Rewritten from one allowlisted primary — not independent enterprise reporting. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →arXiv cs.AI — The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models
Save

A recent study published on arXiv cs.AI investigates the communication bottleneck that arises when language models serialize tree-structured information into natural language and then attempt to extract it. The research proposes a round-trip protocol to empirically assess how much compositional content survives this process for tree-structured expressions.

The protocol involves a generator converting a procedurally generated arithmetic expression into a word problem. A separate extractor then recovers the expression from the word problem. Symbolic equivalence provides an exact oracle for evaluation. The study evaluated all pairwise combinations of sixteen models, generating a communication matrix that differentiates generation quality from extraction quality.

Key Points

  • The communication channel for tree-structured expressions is lossy and asymmetric.
  • Swapping the generator and extractor models can shift accuracy by up to 60.4 points.
  • The best performing pair of models achieved 92.9% accuracy by combining different models for generation and extraction.
  • At least 73.6% of round-trip failures originated during the generation phase.
  • Difficulty in serialization is primarily driven by tree structure, including operator count, depth, and right-branching, rather than by the model family.

Context

According to the authors, language models often serialize structured information into natural language when performing chain-of-thought reasoning or exchanging free-text intermediates. The study's round-trip protocol was designed to quantify the survival rate of tree-structured compositional content through this serialization and deserialization process. By using a generator to create word problems from arithmetic expressions and an extractor to reverse the process, the researchers could precisely measure accuracy using symbolic equivalence.

Why It Matters

This research highlights a fundamental challenge in how language models handle structured data, indicating that the conversion between structured and unstructured formats is not always reliable. Builders and researchers should note that the choice of model for generation versus extraction can significantly impact performance, and that current methods introduce substantial information loss, particularly during the initial serialization step.

What To Do

  • Note that the choice of model for generation and extraction can significantly impact overall accuracy.
  • Consider the implications of a lossy communication channel when designing systems that rely on language models to process or generate structured information.
  • Investigate the structural properties of expressions (operator count, depth, branching) when troubleshooting serialization failures, as these appear to be primary drivers of difficulty.
  • Explore strategies that might mitigate the identified asymmetry and generation-side failures in structured data handling.