XBridge Enables Efficient Cross-Architecture LLM Communication

Multi-agent LLM systems can achieve superior performance when composed of diverse model families, as this reduces redundant reasoning. However, current communication methods either rely on text, losing internal representations, or require identical architectures for latent-level data transfer. A significant challenge, termed the "entity grounding problem," arises when transferring continuous representations across different LLM families, as critical entity identities can be lost. Researchers have introduced XBRIDGE, a novel communication protocol designed to overcome these limitations. XBRIDGE operates without decoding, enabling efficient latent-level communication between heterogeneous LLMs. It employs two key mechanisms: Lexical Anchor Mapping (LAM), which maps sender context tokens to the receiver's vocabulary to provide discrete entity anchors, and a Latent Enrichment Bridge (LEB), allowing the receiver to query the sender's hidden states for contextual details. This approach ensures that contextual signals are grounded to specific entities via the receiver's self-attention. Tested across various model families (Llama, Qwen, Mistral) and benchmarks, XBRIDGE consistently outperformed text-based communication, achieving 11 times lower latency. It also surpassed KV-sharing baselines in same-architecture settings, demonstrating its effectiveness and efficiency with minimal overhead.