XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

2026-08-12

AI Summary

This work addresses the challenge of inefficient semantic communication in heterogeneous large language model (LLM) multi-agent systems, where existing protocols—relying on textual exchange or requiring architectural homogeneity—struggle to convey latent-space semantics and often suffer from entity identity collapse in continuous representations. To overcome these limitations, the authors propose XBridge, which introduces Lexicon Anchor Mapping (LAM) to establish discrete cross-model entity anchors for the first time and designs a lightweight Latent-state Enhancement Bridge (LEB) that enables receivers to directly query senders’ hidden states. Requiring no decoding and only 264M parameters, XBridge is compatible with diverse LLMs such as Llama, Qwen, and Mistral. It significantly outperforms text-based communication—reducing latency by 11×—across seven benchmark tasks and surpasses KV-sharing baselines on six tasks even in homogeneous settings, demonstrating its generality and efficiency.

Abstract

Heterogeneous multi-agent LLM systems, where agents are powered by different model families, can outperform homogeneous configurations by reducing redundant reasoning patterns. Yet existing communication protocols either operate through text, discarding the sender's internal representations, or require architectural homogeneity for latent-level transfer. We identify the entity grounding problem in cross-architecture communication: cross-attention bridges that transfer continuous representations across different LLM families suffer from rare-token compression collapse, where entity identity is lost in the continuous bottleneck (bridge-only F1 ~30%). We propose XBRIDGE, a decode-free communication protocol that addresses this through two mechanisms. Lexical Anchor Mapping (LAM) maps the sender's original context tokens to the receiver's vocabulary, providing discrete entity anchors. A Latent Enrichment Bridge (LEB) lets the receiver query the sender's hidden states for contextual enrichment. The entity anchors ground the bridge's contextual signals to specific entities through the receiver's own self-attention. Across three model families (Llama, Qwen, and Mistral), seven benchmarks, and both communication directions, XBRIDGE outperforms text-based communication on all seven tasks for each model pair while achieving 11x lower latency, and in a same-architecture setting it also exceeds a KV-sharing baseline on six of seven tasks. LEB requires only 264M trainable parameters (3.8% of the receiver), is trained on a small balanced sample set, and adds negligible inference overhead.

Problem

• heterogeneous LLM communication
• entity grounding
• cross-architecture communication
• latent bridge
• rare-token compression collapse

Innovation

• entity grounding
• heterogeneous LLM communication
• latent bridge
• lexical anchor mapping
• latent enrichment