Researchers have developed Hybrid Reasoning Agents (HybridAgents), a new architecture that enables large language models to dynamically shift between two modes of reasoning: latent (internal, efficient) and text-based (externally visible, tool-integrated).
According to the paper, the system bridges a long-standing gap in AI agent design. Modern LLM agents traditionally reason and interact entirely through text, which is transparent but computationally expensive. Latent reasoning offers efficiency but typically lacks native tool access and interpretability. HybridAgents preserves the benefits of both approaches.
The architecture includes two key components. A lightweight HybridLink module enables seamless transitions between latent and text modes during generation. The agent’s continuous working memory maintains lossless hybrid context across multiple conversation turns, allowing it to preserve information regardless of which reasoning mode it is using.
The team developed dedicated post-training algorithms to optimize hybrid reasoning. HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone model to coordinate with HybridLink. HybridRL then scales reasoning performance by learning when to think in latent space versus when to act through text.
Evaluation across 3 model scales and 11 benchmarks—covering mathematics, science, medicine, search, and agentic reasoning tasks—shows consistent improvements. According to the paper, HybridAgent delivered 10.4% better performance compared with frontier LLM agents and competing agentic frameworks. More notably, it achieved 35.5% to 54.7% token reduction, 1.8× to 2.6× faster inference speed, and 2.7× to 3.7× cost savings.
These efficiency gains address a critical bottleneck in deploying reasoning-heavy AI systems at scale. By allowing models to use latent reasoning for internal computation and reserve text for external tool calls and actions, HybridAgents reduce both processing time and the computational resources required per inference. The approach maintains tool integration—critical for real-world agent applications—while making internal reasoning directly inspectable, addressing transparency concerns.
Code and data for the research are available through the paper’s repository.
Key facts
- HybridAgents dynamically switch between latent (internal) and text-based reasoning modes during generation
- The system achieved 35.5%–54.7% token reduction and 1.8×–2.6× faster inference compared to baseline methods
- Cost savings reached 2.7×–3.7× across evaluated tasks
- Performance improved by 10.4% on 11 benchmarks spanning mathematics, science, medicine, search, and agentic reasoning
- The architecture preserves tool-integrated agentic capabilities while making latent reasoning inspectable