Stanford and Google propose latent vector reasoning for AI agents

Researchers from Stanford and Google published Hybrid Reasoning Agents on October 13, 2026, introducing the HybridAgents architecture that alternates text steps with continuous vectors. The hybrid mode improves performance by 10.4%, cuts generated tokens by 35.5–54.7% and speeds up inference 1.8–2.6x across 11 benchmarks.

• Hybrid text-vector reasoning beats text-only and latent-only agents by 10.4%
• Token generation per task drops by 35.5–54.7%
• Inference runs 1.8–2.6x faster with up to 3.7x lower cost
• The lightweight HybridLink module switches modes without full retraining