XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

arxiv:2608.11676

Original source · Paper JSON

LLMs: architectures, context, training · 2026-08-12

Cites (3) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• XBridge : Entity-Grounded Latent Bridge for Heterogeneous LLM Communication
• Abstract
• 1 Introduction
• 2 The Entity Grounding Problem
• 2.1 Entity Fidelity
• 2.2 Entity Grounding Failure
• 2.3 Representation Mismatch
• 3 XBridge : Entity-Grounded Latent Communication
• 3.1 XBridge Overview
• 3.2 Lexical Anchor Mapping (LAM) for Discrete Grounding
• Cross-vocabulary mapping.
• Receiver-native lexical embeddings.
• 3.3 Latent Enrichment Bridge (LEB) for Context Adaptation
• Projections.
• Cross-attention.
• Gated residual.
• 3.4 Entity-Grounded Bridge Integration
• 4 Experiments
• 4.1 Experimental Setup
• Models.
• Tasks and protocol.
• Compared methods.
• 4.2 Main Results
• 4.3 Ablation: Both Anchors and Enrichment Are Necessary
• 4.4 Bridge Mechanism and Role Separation
• 4.5 Beyond Accuracy: Latency, Composability, and Sender Scaling
• 5 Conclusion
• Limitations
• Appendix A LAM Details
• A.1 Cross-Vocabulary Mapping
• A.2 Grounding Format Independence
• Appendix B LEB Design
• B.1 Bridge Architecture Ablations
• Number of bridge modules.
• Sender layer selection.
• Gate initialization.
• Auxiliary losses.
• B.2 Bridge Training Data Composition
• Implication for sender size scaling.
• Appendix C Additional Results
• C.1 Homogeneous Comparison
• C.2 Bridge Contribution per Task
• C.3 Multi-Agent Communication
• Appendix D Additional Analysis
• D.1 Per-Sample Consistency
• D.2 Entity Perturbation Examples
• Appendix E Task Details
• Appendix F Related Work
• Text-based multi-agent communication.
• Latent multi-agent communication.
• Instructions for reporting errors