Skip to content- Sources: primary, discussion
- Summary: HybridAgents uses a lightweight HybridLink module to move between latent and text modes and keeps a continuous working memory across turns. Training uses HybridSFT followed by HybridRL, which learns when to think latently and when to act through text. Across 3 model scales and 11 benchmarks, the authors report a 10.4 percent performance gain, 35.5 to 54.7 percent fewer tokens, 1.8x to 2.6x faster inference, and inspectable latent thoughts. Code and data are released.
- Why it matters: Latent reasoning that keeps tool access could cut agent inference cost, though inspecting latent thoughts is a new problem for agent observability.