pipette
ENEnglish

Selective and Hybrid-State Communication for Latent Multi-Agent Systems

Miles Liao, Aaron Smieszek, Yuhao Ge

Preprint

En palabras de los autores

Efficient communication remains a major bottleneck in large language model multiagent systems. Latent communication reduces repeated natural-language exchange by transferring internal model states between agents, but these transfers remain costly and are typically designed around conventional KV-cache representations. We address this problem from two complementary perspectives. We introduce SelectKV, a boundary-aware method that selectively communicates KV states while protecting inherited context, and HyLaMAS, which extends latent collaboration to hybrid linear-attention models by communicating both KV caches and recurrent states through a shared workspace. Across six benchmarks, SelectKV reduces communicated KV positions and logical KV payload by 9.7-9.9%, with accuracy unchanged on two benchmarks, within one percentage point on three, and six points lower on HumanEval+. Across Qwen3.5 model scales, HyLaMAS reduces decoded output tokens by 53-80% and achieves 1.16-2.78x speedups relative to text-based multi-agent communication. Together, these independently evaluated methods represent two complementary approaches to efficient latent-state communication across conventional Transformer and hybrid model architectures.

Resultado principalLimitación que admiten los autores

Apareció: viernes, 30 de octubre. arXiv. Preprint, todavía sin revisión por pares.

Comentario de los autores: Accepted at the NeurIPS 2026 Workshop on Resource-Aware Agentic AI (RAAAI)