Synapse: Federated Tool Routing via Typed Compendium Artifacts

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Synapse is a novel federated framework designed for LLM-based agents to collaboratively learn tool-usage patterns while preserving data privacy and reducing communication costs. Instead of sharing raw data, model weights, or discrete prompts, Synapse trains a shared global knowledge model using "compendiums"—structured, hierarchical representations encapsulating modular knowledge and contextual information. This tiered architecture involves clients generating local compendiums, edge aggregators refining them, and a central server aggregating and redistributing a global snapshot. Evaluated on reasoning-intensive benchmarks like BBH and GSM8k, Synapse achieves 92–93% accuracy, demonstrating improved tool-usage effectiveness and significantly lower communication overhead compared to weight or prompt-sharing baselines, even with non-IID data distributions.

Key takeaway

For MLOps engineers deploying LLM-based multi-agent systems in privacy-sensitive or bandwidth-constrained environments, Synapse offers a compelling alternative to traditional federated learning. You should consider adopting a compendium-based knowledge exchange approach to achieve high tool-routing accuracy (92-93% on GSM8k) and robust performance with significantly reduced communication overhead. This method allows for collaborative learning of tool-usage patterns across heterogeneous clients while maintaining data privacy, enabling scalable and adaptive AI deployments in regulated domains like healthcare or finance.

Key insights

Synapse enables privacy-preserving federated LLM tool routing by exchanging structured knowledge compendiums instead of model parameters.

Principles

• Compendiums offer richer expressivity than prompt-sharing.
• Low communication overhead can match centralized performance.
• Fixed noise ensures routing convergence despite privacy.

Method

Clients curate local compendiums, edge aggregators deduplicate/summarize, a central server aggregates into a global snapshot. This snapshot is redistributed, enabling tool selection via vector search, LLM reranking (llama-3.1-8b-instruct), and LLM planning.

In practice

• Generate usage scenarios locally with domain-specific data.
• Apply adaptive text masking and Laplace noise for privacy.
• Use LLM rerankers (e.g., llama-3.1-8b-instruct) for tool selection.

Topics

• Federated Learning
• LLM Tool Routing
• Retrieval-Augmented Generation
• Differential Privacy
• Multi-Agent Systems
• Knowledge Representation

Best for: AI Architect, Research Scientist, AI Scientist, MLOps Engineer, AI Security Engineer

Related on AIssential

• Echelon: Auditable Aggregate-Only Language-Model Adaptation Across Privacy Boundaries — Echelon enables secure, auditable cross-organizational LLM adaptation by exchanging only aggregated deltas, preserving device-level privacy.
• AgentGate: A Lightweight Structured Routing Engine for the Internet of Agents — AgentGate enables efficient, privacy-aware AI agent routing via a two-stage, structured decision process for compact models.
• Are Tools All We Need? Unveiling the Tool-Use Tax in LLM Agents — Tool-use in LLM agents incurs a "tool-use tax" that can negate performance gains, especially with semantic noise.
• Privacy-Preserving LLMs Routing — PPRoute enables privacy-preserving LLM routing via MPC-friendly operations, multi-stage training, and an efficient unsorted Top-k algorithm.
• SplitFT: An Adaptive Federated Split Learning System For LLMs Fine-Tuning — SplitFT adaptively optimizes federated split learning for LLM fine-tuning, addressing heterogeneity and communication overhead.
• MetaMoE: Diversity-Aware Proxy Selection for Privacy-Preserving Mixture-of-Experts Unification — MetaMoE unifies distributed MoE models using diversity-aware public proxy data to preserve privacy and enhance expert coordination.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.