This work proposes LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration, which achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.
Abstract
LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.
XKV is introduced, which lifts all three restrictions on multi-agent LLM answering: learned-query attention pools both caches; self-attention over receiver-aligned layer tokens, with a learned layer map reconciling different depths, mixes the pooled summaries into a compact joint memory.
Ji-Yao Liu, Qi Zhang, Yaoyi Jia et al.· 3 citations
CacheBack is a simple, robust, training-free instance of receiver conditioning based on the sender's attention weights that improves accuracy and reduces median task-completion latency on FanOutQA and shows comparable improvements across model families that span dense Transformers, Mamba-attention hybrids, and sliding-...
M. Rossi, Prajwal Raghunath, Hao Xuan et al.· 0 citations
Large language model-based multi-agent systems improve complex problem solving through collaboration, while latent communication directly transmits model internal states to avoid the high inference costs of natural language. However, existing KV-based latent communication methods prioritize sender-side state fidelity,...
Dong-Sen Zhang, Pei-Pei Li, Ze-Kun Li et al.· 0 citations
Multi-agent systems for automated program repair communicate through text tokens, requiring each agent to decode its internal state into discrete symbols and the next agent to re-encode them. This conversion loses information and incurs token overhead that grows quadratically with the agent count. We develop Latent Age...
Rui Jiao, Yue Zhang, Zhe-Xuan Feng et al.· Proceedings of the 41st IEEE...· 0 citations
Large language models (LLMs) have demonstrated remarkable capabilities in solving complex tasks in coding, mathematical logic, and iterative reasoning, especially by coordinating effectively within multi-agent systems (MAS). However, language agents traditionally communicate via text exchange, which introduces a bottle...
This work proposes AgentSpec, a speculative decoding algorithm that addresses the limitations of existing methods for LLM agents and incorporates structure-isolated drafting that constrains speculation to semantically coherent segments of the agent workflow, reducing the drafts of irrelevant semantic paths and achievin...