arXiv:2607.14119v1 Announce Type: new Abstract: Multi-agent LLM systems commonly decompose complex tasks into specialized roles. However, this modularity introduces a representational risk: when intermediate agents transform text across linguistic registers, they can systematically compress the semantic distinctions needed for accurate downstream decisions. We term this phenomenon semantic register compression and characterize it as an observable failure mode in multi-agent cascades. Using a three-agent pipeline (Collector-Evaluator-Decider), we quantify compression via inter-label separation in sentence-transformer embedding space. Across political fact-checking (LIAR), sentiment analysis (SST-5), and medical triage (Triagegeist), critical evaluation consistently reduces label separability by 41.7% at the Evaluator stage, while identity passthrough preserves it nearly fully. Five architectural variants causally isolate oriented semantic transformation as the primary driver. A credibility-seeking variant produces minimal geometric compression yet shifts outputs toward mostly-true, demonstrating that transformation valence controls the direction of distributional collapse independently of compression magnitude. Compression generalizes across the three domains with varying intensity: 41.7% in fact-checking, 27.2% in sentiment, and 20.0% in triage. Prompt-level regression explains 78% of the variance, with operational constraints associated with lower compression. These results demonstrate that semantic register compression is a measurable and generalizable phenomenon in multi-agent LLM systems, with implications for safety evaluation in high-stakes domains.
Anthropic cuts internet access for AI evaluations to prevent unintended actions