DIALECT: Continuous-Space Communication Within and Across Language Models

Multi-agent systems (MAS) enable large language models (LLMs) to collaborate on complex tasks by combining complementary capabilities and exchanging intermediate reasoning. MAS primarily communicate through natural language, which introduces latency through message generation and limits the information passed to the receiver by selecting discrete tokens from distributions over possible continuations. Continuous representations offer an alternative communication medium, but differences in tokenization, embedding dimensionality, and representational structure pose challenges to cross-model communication. We introduce Distributional Inter-Agent Latent Exchange via Cross-model Translation (DIALECT), a framework for continuous-space communication between agents using the same or different LLMs. The source agent constructs continuous thoughts as stochastic mixtures of token embeddings, retaining contributions from multiple continuations at each step. For same-model communication, thoughts transfer directly; for cross-model communication, a bridge trained in two stages maps them into the target's input embedding space, with both LLMs frozen. We evaluate DIALECT in a two-agent planning-and-solving setting on 13 benchmarks across mathematical reasoning, knowledge question answering, and code generation across four models from three families, spanning 3B–8B. DIALECT significantly outperforms single-agent targets, same-model baselines, and cross-model baselines. Compared to text transfer constrained to equal source generation steps, it achieves significantly higher accuracy with lower latency. Against full-text communication, it maintains statistically equivalent accuracy in all model pairs, while delivering a median speedup of 2.02$\times$. These results support stochastic soft-token generation and a learned cross-model bridge as a practical approach to continuous-space communication within and across LLMs.