Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synapse Compendium Aware Federated Knowledge Exchange for Tool Routed LLMs

Abhijit Chakraborty, Sandipan De, Yash Shah, Chahana Dahal, Vivek Gupta

TL;DR

Synapse addresses privacy and heterogeneity in federated tool routing for LLMs by exchanging structured compendiums rather than model weights or prompts. The framework uses a three-tier federation, hierarchical compendium aggregation, embedding-based retrieval, LLM reranking, and TextGrad prompt optimization to dynamically route queries to tools. Empirical results on BBH and GSM8k show improved routing accuracy and robustness with reduced communication overhead compared with prompt-sharing and weight-sharing baselines, even under non-IID data. The work demonstrates a practical approach to privacy-preserving, scalable knowledge exchange in multi-agent LLM ecosystems.

Abstract

Collaborative learning among LLM-based agents under federated learning faces challenges, including communication costs, heterogeneity in data, and tool-usage, limiting their effectiveness. We introduce Synapse, a framework that trains a shared global knowledge model of tool-usage behavior. Client agents with fixed LLMs learn tool-usage patterns locally, and transmit artifacts for federated aggregation through coordinators. A global tool compendium is updated and redistributed, enabling convergence toward stable tool selection. Synapse uses templated representations, embedding retrieval with LLM reranking, and adaptive masking to maintain utility while limiting information leakage. The framework supports heterogeneous data and quantifies performance improvements. Results show that Synapse improves tool-usage effectiveness and reduces communication overhead compared with weight or prompt-sharing approaches in multi-agent LLM systems.

Synapse Compendium Aware Federated Knowledge Exchange for Tool Routed LLMs

TL;DR

Synapse addresses privacy and heterogeneity in federated tool routing for LLMs by exchanging structured compendiums rather than model weights or prompts. The framework uses a three-tier federation, hierarchical compendium aggregation, embedding-based retrieval, LLM reranking, and TextGrad prompt optimization to dynamically route queries to tools. Empirical results on BBH and GSM8k show improved routing accuracy and robustness with reduced communication overhead compared with prompt-sharing and weight-sharing baselines, even under non-IID data. The work demonstrates a practical approach to privacy-preserving, scalable knowledge exchange in multi-agent LLM ecosystems.

Abstract

Collaborative learning among LLM-based agents under federated learning faces challenges, including communication costs, heterogeneity in data, and tool-usage, limiting their effectiveness. We introduce Synapse, a framework that trains a shared global knowledge model of tool-usage behavior. Client agents with fixed LLMs learn tool-usage patterns locally, and transmit artifacts for federated aggregation through coordinators. A global tool compendium is updated and redistributed, enabling convergence toward stable tool selection. Synapse uses templated representations, embedding retrieval with LLM reranking, and adaptive masking to maintain utility while limiting information leakage. The framework supports heterogeneous data and quantifies performance improvements. Results show that Synapse improves tool-usage effectiveness and reduces communication overhead compared with weight or prompt-sharing approaches in multi-agent LLM systems.
Paper Structure (37 sections, 3 theorems, 15 equations, 13 figures, 5 tables, 1 algorithm)

This paper contains 37 sections, 3 theorems, 15 equations, 13 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Synapse Framework
  4. Generation and Curation of Usage Scenarios on Clients:
  5. Synapse Workflow:
  6. Evaluation of Synapse:
  7. Experiments and Analysis
  8. Evaluation Results
  9. Performance Analysis Across Client Setups
  10. Federated Learning Setup Comparison (IID vs. non-IID splits)
  11. Scalability Tests
  12. Privacy Analysis
  13. Ablation Study
  14. Tool Convergence in Federated Routing: Stability under Stochastic Artifact Perturbations
  15. Impact of Distinct Failure Modes on System Accuracy and Stability
  16. ...and 22 more sections

Key Result

Theorem 1.1

Let $s$ be a usage scenario text and $\tilde{s} = \mathsf{PrivTrans}(s)$ its privacy-transformed version. Let $e(\cdot)$ denote the embedding function. Under privacy budget $\varepsilon$, there exists a distortion bound $\delta(\varepsilon)$ such that: Consequently, the similarity score deviation for any query $q$ is bounded by:

Figures (13)

  • Figure 1: Tiered architecture of Synapse showing client, edge, and central aggregation of compendiums for federated tool routing.
  • Figure 2: Shows an example compendium for a mathematical word problem solver. The modular structure maintains semantic coherence for retrieval. Yellow highlights indicate newly added tool scenarios, while blue highlights indicate updated scenarios learned over time.
  • Figure 3: Inference Retrieval and Routing Pipeline in the FL system within privacy as described in algorithm \ref{['alg:synapse']}.
  • Figure 4: Tool-routing accuracy on GSM8K under federated settings, comparing SYNAPSE with prompt-sharing and retrieval-based baselines.
  • Figure 5: Fairness and heterogeneity analysis across federated rounds: (a) accuracy spread and (b) macro accuracy.
  • ...and 8 more figures

Theorems & Definitions (6)

  • Theorem 1.1
  • proof
  • Theorem 1.2
  • proof
  • Theorem 1.3
  • proof