AI agents' text communication destroys 88% of internal features, study finds

A new paper on arXiv (2607.14103) by Markus Wenzel challenges the assumption that clear-text message passing is sufficient for multi-agent LLM systems. The researchers hypothesized that LLMs have a world model exceeding textual expressibility. They constructed three communication channels—text, dense latent, and sparse latent—and used Sparse Autoencoder (SAE) feature analysis to measure information retention. Results showed that the SAE-sparse channel retained 99.4% probe accuracy at 28-fold compression over the dense-latent channel, compared to just 80.4% for text. In cross-architecture tests (Llama ↔ Mistral) using Procrustes alignment, they achieved 92% top-1 retrieval. But a text round-trip feature survival analysis revealed that serialization destroyed 88% of SAE features, replacing them with a different set.

Despite the apparent loss, the paper arrives at a surprising negative conclusion: the destroyed features mostly encode surface form, not task-relevant semantics. When tested on cross-lingual concept tasks, the latent channel matched—but never exceeded—the text channel. Augmenting text with latent features provided no benefit, and performance penalties from alignment were only 3–10 percentage points. This suggests that current benchmarks may be too shallow to leverage latent communication's potential. The authors call for deeper tasks that elicit complex concepts and better analysis frameworks before dismissing text-based inter-agent communication.