New Agent Architectures Push Reasoning Into Latent Space — and Let Robots Teach Themselves
Two research efforts point the same direction: HybridAgents reasons without writing every step as text, while SEAR runs the first large-scale study of robot agents that improve with practice.
Two research releases this week quietly challenge the dominant assumption that AI reasoning has to happen in written text. The first, HybridAgents, comes from researchers at Stanford, UIUC, Meta, and Together AI. As @0xdominus reported, it lets models reason in latent space instead of spelling out every intermediate step as language. The reported payoff is substantial: roughly 10% higher scores on many tests, half the token usage, and twice the speed.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.