Synapse: Federated Tool Routing via Typed Compendium Artifacts

Researchers introduce Synapse, a federated learning framework using typed artifacts that enables heterogeneous language models to collaborate without sharing weights or data. The system enables cross-architectural model transfer with minimal performance loss while maintaining formal privacy guarantees and schema-aware merging capabilities.

Synapse addresses a fundamental limitation in federated learning: existing approaches treat collaboration units—weights, prompts, or raw examples—as untyped objects, preventing well-defined privacy operations and cross-model compatibility. The research proposes typed federated artifacts with validated schemas, enabling per-field differential privacy and structured conflict resolution. This matters because federated learning across diverse models has remained largely theoretical; weight-sharing federation requires architectural matching, eliminating heterogeneity benefits.

The framework emerges from growing demand for privacy-preserving AI systems and the proliferation of open-source language models with incompatible architectures. Traditional federated approaches either leak gradient information or discard structural information to maintain privacy. Synapse solves this by treating the compendium—a shared artifact repository—as the collaboration unit rather than model weights, allowing frozen, heterogeneous LLMs to participate without architectural alignment requirements.

The practical implications extend beyond academic interest. Organizations deploying multiple LLM providers face a choice between privacy-leaking centralization and capability-limiting isolation. Synapse's demonstrated transfer across LLaMA 3.1, LLaMA 3.2, Mistral, and GPT-4o families with approximately 2-point loss suggests enterprise-grade federated inference becomes viable. This could reshape how organizations build AI systems, enabling tool routing and knowledge sharing without exposing proprietary model weights or training data.

Future developments should examine scaling to larger model families, real-world deployment scenarios, and whether formal DP guarantees hold under adaptive adversaries. The routing-stability characterization across five distributions provides empirical grounding, though the contraction premise failure case warrants deeper investigation into failure mode mitigation.

• →Synapse enables federated collaboration between heterogeneous LLMs without weight-sharing or architectural matching requirements.
• →Typed artifacts with schema validation provide formal differential privacy guarantees per-field rather than heuristic approximations.
• →Single compendium transfers across four LLM families with approximately 2-point performance loss, a capability impossible in traditional weight-sharing federation.
• →The framework preserves privacy while enabling cross-model tool routing and knowledge transfer in frozen, decentralized LLM environments.
• →Field-wise conflict resolution and conditional retrieval distortion enable structured collaboration without gradient leakage or data sharing.