PROTOCOL FOR DYNAMIC ADAPTATION OF MESSAGE INFORMATION DENSITY IN MULTI-AGENT SYSTEMS BASED ON LARGE LANGUAGE MODELS
DOI:
https://doi.org/10.34185/1991-7848.itmm.2026.01.088Keywords:
multi-agent systems, large language models, adaptive compression, inter-agent communication, context load managementAbstract
The increasing complexity of tasks solved by multi-agent systems based on large language models imposes higher demands on the efficiency of inter-agent communication. Existing approaches to context compression do not take into account the current state of the receiving agent, which leads to output quality degradation, increased latency, and quadratic growth of token traffic when scaling the system. This work proposes the Receiver-Load-Aware Compression Protocol (RLACP), an adaptive inter-agent communication protocol in which the level of message compression is dynamically determined by the current cognitive load of the receiving agent. This load is formalized as a composite metric consisting of four components: number of active tasks, semantic uncertainty of outputs, response latency, and context saturation level. Depending on the value of this metric, the compressing agent applies one of four compression modes — ranging from preserving full text to complete abandonment of natural language in favor of structured key-value pairs.
References
Liu N. F. et al. Lost in the Middle: How Language Models Use Long Contexts. URL: https://arxiv.org/abs/2307.03172
Jiang H. et al. LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. URL: https://arxiv.org/abs/2310.05736
Chevalier A. et al. Adapting Language Models to Compress Contexts. URL: https://arxiv.org/abs/2305.14788
Filippova, K., & Strube, M. (2008). Dependency Tree Based Sentence Compression. Proceedings of the Fifth International Natural Language Generation Conference (INLG'08), pp. 25–32. Association for Computational Linguistics. URL: https://aclanthology.org/W08-1105/
Radev, D., Jing, H., Stys, M., & Tam, D. (2004). Centroid-Based Summarization of Multiple Documents: Sentence Extraction, Utility-Based Evaluation, and User Studies. Proceedings of the NAACL Workshop on Automatic Summarization.
URL: https://dl.acm.org/doi/10.3115/1117575.1117578
LLM-TKIE: Large Language Model Driven Transferable Key Information Extraction Mechanism. (2025). Scientific Reports. URL: https://www.nature.com/articles/s41598-025-15627-z
