Enabling Agents to Communicate Entirely in Latent Space

arxiv:2511.09149

Original source · Paper JSON

LLMs: architectures, context, training · 2025-11-12

Cites (6) · 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

• Enabling Agents to Communicate Entirely in Latent Space
• Abstract
• 1 Introduction
• 2 Related Work
• Latent Reasoning in LLMs.
• Multi-agent Communication.
• 3 Interlat
• 3.1 Latent Communication
• 3.2 Training Procedure
• Conditional thought separation.
• Plan-aligned regulation.
• Curriculum Learning.
• 4 Information Compression
• Compression via Latent-Space Reasoning.
• Training Objective.
• 5 Experiments
• Implementation Details.
• Baselines and variants in Interlat.
• 5.1 Main Results
• Latent Communication Prompts Exploration.
• Semantics and Learning Dynamics.
• Generalization to Symbolic Reasoning.
• 5.2 Compression Analysis
• Compression Performance.
• Why compression is effective.
• 5.3 Ablation Studies
• 5.4 Generalization to Complex Topologies.
• 6 Conclusion
• 7 Limitations
• 8 Ethical Considerations
• 9 Acknowledgments
• Appendix A LLM usage
• Appendix B Compression Loss
• Setup.
• (1) Actor cross-entropy utility.
• (2) Uncertainty-weighted agreement.
• (3) Latent direction alignment.
• Overall objective.
• B.1 Baselines and settings in Interlat.
• Appendix C Benchmark
• C.1 Alfworld
• Appendix D Ablations and Step Analysis
• Effect of curriculum learning.
• Step count versus performance.
• Appendix E Compression Result
• Appendix F Latent Parallelism Analysis
• Appendix G Qualitative Analysis of Latent Communication via PCA
• Appendix H Training Template
• Instructions for reporting errors