HybridAgents: Latent Thinking, Text Acting
Original: Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
Why This Matters
Merging latent reasoning with tool-use could redefine efficiency benchmarks for agentic AI.
Stanford and UIUC researchers propose HybridAgents, LLM agents that reason in latent space but act through text, achieving 10.4% performance gains, up to 54.7% fewer tokens, and 3.7x cost savings across 11 benchmarks.
A team from Stanford, UIUC, and UCLA has released HybridAgents, a framework that splits the cognitive labor of LLM agents: internal reasoning happens in latent (continuous) space, while external actions and tool calls remain text-based. The core architectural addition is HybridLink, a lightweight module that handles transitions between the two modes within a single generation pass. Continuous working memory preserves context across turns without information loss.
Training uses two dedicated stages. HybridSFT first stabilizes tool-integrated hybrid generation by teaching the base model to coordinate with HybridLink. HybridRL then trains the agent to decide dynamically when to think in latent space versus when to produce text—a policy the authors argue is learned rather than hardcoded.
Evaluated at three model scales across 11 benchmarks spanning math, science, medicine, web search, and multi-step agentic tasks, HybridAgent outperforms both text-only and latent-only frontier systems by an average of 10.4%. Token usage drops 35.5%–54.7%, inference speeds up 1.8x–2.6x, and cost falls 2.7x–3.7x. The authors also note that latent thoughts remain inspectable, addressing a common transparency concern with non-text reasoning. Code and data are publicly available.