Search papers, labs, and topics across Lattice.
This paper introduces XBRIDGE, a novel decode-free communication protocol designed to enhance interaction between heterogeneous multi-agent LLM systems by addressing the entity grounding problem. By employing Lexical Anchor Mapping (LAM) and a Latent Enrichment Bridge (LEB), XBRIDGE effectively maintains entity identity and contextual relevance across different model families, significantly outperforming traditional text-based communication methods. The results demonstrate that XBRIDGE not only achieves superior performance across seven benchmarks with reduced latency but also requires minimal additional parameters, making it a highly efficient solution for cross-architecture communication.
XBRIDGE reduces communication latency by 11x while ensuring that heterogeneous LLMs maintain entity identity and contextual relevance.
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.
