Summary
Two Minute Papers host Dr. Károly Zsolnai-Fehér covers a research paper proposing a fundamentally different approach to multi-agent coordination: instead of agents communicating through natural language, they pass raw latent state vectors — undecoded neural activations — directly between models. The technique, called cross-agent latent state transfer, eliminates the encode-to-text and text-to-decode overhead that makes standard agent pipelines slow and token-intensive.
The results on competition-level math problems are significant. A three-agent pipeline using latent communication improves accuracy from 73% to 86% compared to text-based equivalents, while reducing token usage by 75%. Total training cost for the setup was approximately four dollars. The models tested are sub-10-billion-parameter systems — not frontier models — yet the approach brings them into competitive range with much larger systems on structured reasoning tasks. The video also highlights a potential new scaling behavior: adding more rounds of latent refinement continues to improve results up to an optimal latent thought length of around 80 steps.
Zsolnai-Fehér addresses the key methodological question directly: is the gain from the architecture or from being distilled by a high-quality teacher model? The paper includes a controlled comparison giving the same teacher to multiple architectures, confirming the latent communication structure itself drives the improvement. Limitations are clearly stated — results are on smaller models and scalability to frontier systems is unknown — and code and model weights are publicly available.
📺 Source: Two Minute Papers · Published June 19, 2026
🏷️ Format: Deep Dive
