AI researchers introduce internalized multi‑agent debate and high‑bandwidth latent reasoning

Researchers have presented two new approaches to improve large language model (LLM) reasoning. One method, called internalized multi‑agent debate, distills the outcomes of multi‑agent discussions into a single model through a two‑stage fine‑tuning pipeline. The resulting models achieve comparable or better performance than explicit debate while using up to 93% fewer tokens. Analysis shows the internalized model creates agent‑specific subspaces in activation space, enabling targeted steering to suppress malicious behavior with minimal impact on overall performance.

A separate framework, NF‑CoT (Normalizing‑Flow Chain‑of‑Thought), replaces the serial, text‑based chain‑of‑thought process with continuous latent reasoning using normalizing flows. This approach preserves the autoregressive strengths of standard LLMs while providing high‑bandwidth reasoning, leading to boosted performance and efficiency on tasks such as code generation.