Researchers demonstrated that AI agents communicating through raw numerical states (called cross-agent latent state transfer) instead of English text achieve 86% success on competition-level math problems, down from 73% using language. This cuts token consumption by 75% and trains for four dollars, suggesting that the bottleneck is not model size but how agents exchange information—opening a path to cheaper, smaller AI systems that rival expensive larger ones.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Researchers found that when AI agents exchange internal numerical states directly instead of writing sentences to each other, they solve math problems at 86% success (up from 73%), use 75% fewer tokens, and train for just four dollars on sub-10 billion parameter models.
Why it matters
The discovery suggests that forcing AI agents to communicate in natural language creates a massive bottleneck—each agent must encode its thoughts into sentences, then the next agent must decode them back. Removing that translation step means smaller, cheaper models can match the performance of much larger systems, which may reshape the economics of AI deployment.
What to watch
The approach works on smaller models tested so far; whether it scales to larger ones remains unknown. Latent communication plateaus at around 80 steps, though agents already solve Olympiad-level math problems within that limit.
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
