Learning Communication between Language Models through Dense Vectors

ICLR 2026 Conference Submission11951 Authors

Keywords: Large language models, Neural Architecture, Multi-model collaboration

Abstract: Communication between language models plays a crucial role in the inference process of large language models (LLMs), occurring both iteratively within a single model for multi-step reasoning (auto-regression) and interactively across multiple models for collaborative intelligence. While current systems primarily facilitate such communication through natural language, this paper proposes a novel paradigm of using continuous dense vector in continuous space. Our approach eliminates the unnecessary embedding and de-embedding steps when LLM interact with another, enabling more efficient information transfer, fully differentiable optimization path, and exploration of capabilities beyond human heuristics. We place such stripped LLMs as vertexes and optimizable seq2seq modules as edges to construct LMNet, a directed graph with similar structure as MLPs, and performs end-to-end gradient-descent for efficient optimization. As two exemplar applications, we show the proposed method can effectively improve LLM's general intelligence, and customizing LLM with limited data. We also provide detailed discussion and analysis about learning communication through dense vectors.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 11951