arXiv cs.CL· Wooseong Yang, Wei-Chieh Huang, Weizhi Zhang, Yu Wang, Philip S. Yu, Junhyun Lee·· 11 小时前AI 评分39
XBridge：面向异构 LLM 通信的实体锚定潜在桥接方案
XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication
AI 导读
XBridge 是一种免解码通信协议，通过词法锚定映射（LAM）和潜在增强桥（LEB）解决跨架构 LLM 通信中的实体锚定问题。在 Llama、Qwen、Mistral 三个模型族、七个基准和双向通信上，XBridge 全部七项任务均优于文本通信，延迟降低 11 倍。LEB 仅需 264M 可训练参数（占接收方 3.8%），推理开销可忽略，该工作已被 NeurIPS 2026 接收。
正文 · 原文
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.
| Comments: | Accepted at NeurIPS 2026. 18 pages, 4 figures |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2608.11676 [cs.AI] |
| (or arXiv:2608.11676v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.11676 arXiv-issued DOI via DataCite |
Submission history
From: Wooseong Yang [view email]
[v1]
Wed, 12 Aug 2026 05:32:33 UTC (211 KB)
[v2]
Mon, 28 Sep 2026 18:59:45 UTC (196 KB)
来源：arXiv cs.CL · arxiv.org