Model-to-model text link is lossy: best pair hits 92.9%

The study frames natural language as a channel that models must squeeze structured information through, and asks how much tree-shaped content survives the trip. To test this, the authors built a round-trip protocol: a generator turns a procedurally generated arithmetic expression into a word problem, a separate extractor recovers the expression from the word problem alone, and symbolic equivalence acts as an exact oracle. Running every pairwise combination of sixteen models produces a communication matrix whose marginals separate how well a model writes from how well it reads.

The results point in three directions. The channel is lossy and asymmetric, since the identity of the writer and the reader matters more than the model family. The difficulty tracks the shape of the expression — operator count, depth, right-branching — rather than which lab built the model, and most failures begin on the generation side. And the channel is trainable: a modest set of fine-tuning examples sharing the evaluation's operators and tree shapes lifts every open-weight model above untrained Gemini-3.1-Pro, an upper bound under matched semantics, while a disjoint-domain regime with new operators and vocabulary also raises every open-weight model, confirming the gain is not an artifact of matched semantics.

For anyone wiring models together in multi-step or multi-agent setups, the practical reading is that who writes and who reads is a design choice worth testing rather than assuming away. The remaining gap to the frontier suggests the bottleneck is not fully closed by the fine-tuning described, and the authors identify tree-structured expression serialization as a primary limiting factor when models communicate hierarchical structure through natural language.