RecursiveMAS cuts multi-agent latency by using embeddings only, delivering 2.4x speed and 75% fewer tokens

Agents “talk” through embeddings, not intermediate text—fast and cheaper

Researchers from UIUC and Stanford propose RecursiveMAS, a multi-agent framework that replaces text-to-text communication with latent embedding passing. Instead of generating reasoning tokens at every step, agents loop continuous representations through RecursiveLink modules and only output text at the end. Tests across nine benchmarks show up to 2.4x faster inference, 75% token reduction by round three, and an 8.3% accuracy gain, with far cheaper training than full fine-tuning.