XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

August 13, 2026 | Source: arXiv AI

Tags: multi-agent systems, LLM communication, Llama, Qwen, Mistral, latent representations

XBridge lets Llama, Qwen, and Mistral agents communicate via latent representations rather than text, outperforming text-based messaging on all 7 benchmarks for each model pair while achieving 11x lower latency — using only 264M trainable parameters (3.8% of the receiver model).

Details

Multi-agent systems running different model families have to pass information through text, losing the sender model's internal representations. Existing latent transfer approaches require architectural homogeneity. XBridge solves this with two mechanisms: Lexical Anchor Mapping (LAM) maps the sender's context tokens to the receiver's vocabulary, providing discrete entity anchors that prevent identity collapse; and a Latent Enrichment Bridge (LEB) lets the receiver query the sender's hidden states for richer contextual signals. The entity grounding is the key insight. Prior cross-architecture bridges suffered from what the authors call rare-token compression collapse — continuous representations in the bottleneck lose specific entity identity (achieving only ~30% F1). By anchoring the latent transfer to discrete vocabulary tokens, XBridge maintains entity fidelity. Results across three model families (Llama, Qwen, Mistral), seven benchmarks, and both communication directions show XBRIDGE outperforms text-based communication on all tasks while achieving 11x lower latency. The LEB requires only 264M trainable parameters — 3.8% of a typical receiver model — and adds negligible inference overhead. This opens practical multi-model multi-agent systems where each agent can run a different model family.