Reproduce “Let Models Speak Ciphers: Multiagent Debate through Embeddings” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chau Pham, Boyi Liu, Yingxiang Yang, Zhengyu Chen, Tianyi Liu, Jianbo Yuan, Bryan A. Plummer, Zhaoran Wang and 1 more

Researchers at Boston University and ByteDance Inc. introduced CIPHER, a communication protocol that enables multi-agent Large Language Model (LLM) debates using embedding representations instead of natural language. This method mitigates information loss from discrete token sampling, improving reasoning performance for various open-source LLMs by 1.0% to 5.0% over natural language debate baselines and making debate effective for smaller models.

Key insights

  • Communicating through dense embedding vectors effectively mitigates information loss inherent in natural language's discrete token sampling, leading to more robust inter-LLM exchange.
  • CIPHER enables smaller and less powerful open-source LLMs to consistently benefit from multi-agent debate, broadening the applicability of reasoning enhancement techniques.
  • High-temperature agents in CIPHER can convey valuable, less-confident information through their broader probability distributions, which is effectively utilized by other agents.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.