SILP : A SEMANTIC INTERLINGUA LAYER PROTOCOL FOR CROSS-MODEL AGENT COMMUNICATION

Large language models (LLMs) increasingly communicate with one another through text-based agentprotocols such as MCP and A2A, yet the encoding of semantic payloads in these protocols remainsunspecified. We introduce SILP (Semantic Interlingua Layer Protocol), a black-box, text-interfacepayload codec that compiles a coarse-grained action-slot intermediate representation (IR) into multiplepluggable surface frontends—code-like function-call syntax, pure JSON, natural language, and MLcompressed text—each designed to exploit the shared training priors of contemporary LLMs. Weempirically test the central hypothesis that shared syntactic priors (code-format familiarity frompretraining) contribute more to cross-model comprehension than shared vocabulary priors (word-levelsemantics). Across four main phases plus a smoke-test subphase spanning 5 frontends, 6 tokenizers, 5API-reported model endpoints, and 27 semantically diverse test cases (3,159 total model invocations;counting methodology defined in §3.7), we find: (1) the code frontend achieves the highest averagepass rate (85.9%, statistically tied with JSON at 84.4% and natural language at 83.0%, McNemarp > 0.59) with the lowest cross-tokenizer variance (range 2–10 tokens); (2) ablation reveals thatremoving vocabulary semantics collapses pass rates to 1.2% (skeleton condition), while word-ordershuffling reduces them to 55.6% (shuffled condition), indicating both perturbations caused significantperformance drops but, contrary to the original hypothesis, semantic-vocabulary removal caused alarger performance drop than word-order disruption (85.2 vs. 30.8 percentage-point performancedrops under two distinct perturbations, a 2.8× descriptive ratio); (3) under the tested LLMLingua2 configuration (rate=0.5), the code frontend achieves a 60.0-percentage-point higher pass rate;and (4) in multi-turn sessions of up to 15 rounds, we detect no statistically significant evidence oferror propagation or pass-rate degradation (Fisher exact test, p ≥ 0.10; Benjamini–Hochberg FDRcorrected), although the experiment is underpowered for small deep-context effects. We release theimplemented codec, benchmark harness, protocol specification, and all raw data as open-source athttps://github.com/Juwan-Hwang/polyglotir.