Systems that combine several large language models often rely on a single shared context to answer many different questions. According to the abstract for LatComm, when that context is passed between models as plain text, the receiving model must process every token again for each new question. That repeated processing becomes a significant bottleneck in multi-model pipelines.
The paper, posted on arXiv, introduces LatComm as a way to compress the context for latent communication between LLMs. Instead of sending the full text, the method encodes the context into a compact latent representation that can be reused across queries. This reduces the computational cost on the receiving side.
The abstract does not provide implementation details or experimental results, so the reported benefits are limited to the described design goal. Still, the approach targets a real inefficiency in current multi-LLM architectures, where context reuse is common but token reprocessing is wasteful.