Reproduce “LatComm: Reusable Context Compression for Latent Communication between Large Language Models” on OpenResearch

About this paper

Huawei Lin, Jinting Liu, Tan Yu, Weijie Zhao

LatComm reduces repeated document reading by encoding a context once into a compact packet with a frozen language model, then letting a different frozen receiver read it for each question. On QuALITY, a Qwen3.5-0.8B receiver with a Qwen3.5-4B encoder scores 65.43% Task Avg, versus 61.98% for BGE-Gemma2. It cuts receiver input from 5,915 to 103 tokens per question, but evidence is limited to English multiple-choice QA.

Key insights

• Separating one-time context encoding from question-specific reading enables the same packet to support multiple future questions.
• The packet carries useful context: on the main task suite, accuracy was 65.39% with the correct packet, versus 45.59% with a packet from another context and 39.60% with no packet.
• The approach worked across tested model sizes and model families, though each model pair requires its own trained interface and packet generation.

Summary by alphaXiv.