Tree Expression Serialization Round-Trip Study
A study examining information loss when language models serialize tree-structured expressions into natural language and then reconstruct them. It proposes a round-trip protocol and evaluates pairwise combinations of 16 models to quantify the reliability of natural language as a carrier for structured information transmission.
Timeline (last 90 days)
The study proposes a 'round-trip' experimental protocol comprising three components—generator, extractor, and symbolic equivalence判定—to quantify the communication bottleneck when language models serialize structured information.
The study uses tree-structured arithmetic expressions as controlled probes to investigate the information loss when natural language conveys compositional structural information.
The study conducted a full pairwise evaluation of all two-way combinations among sixteen models, covering cross-model encoder-decoder pairings.
The full pairwise design can disentangle two dimensions of a model: 'language expression capability' and 'structural comprehension capability'.
Arithmetic expressions are generated programmatically at random, enabling systematic control over the complexity, nesting depth, and structural form of the expressions.
When the generator and extractor are different models, the two parties' encoding conventions for natural language are not necessarily consistent, which may amplify communication loss.
All Facts (6)
The study proposes a 'round-trip' experimental protocol comprising three components—generator, extractor, and symbolic equivalence判定—to quantify the communication bottleneck when language models serialize structured information.
50%UnverifiedThe study uses tree-structured arithmetic expressions as controlled probes to investigate the information loss when natural language conveys compositional structural information.
50%UnverifiedThe study conducted a full pairwise evaluation of all two-way combinations among sixteen models, covering cross-model encoder-decoder pairings.
50%UnverifiedThe full pairwise design can disentangle two dimensions of a model: 'language expression capability' and 'structural comprehension capability'.
50%UnverifiedArithmetic expressions are generated programmatically at random, enabling systematic control over the complexity, nesting depth, and structural form of the expressions.
50%UnverifiedWhen the generator and extractor are different models, the two parties' encoding conventions for natural language are not necessarily consistent, which may amplify communication loss.
50%