HybridAgent architecture optimizes LLM reasoning via latent space

Researchers have introduced Hybrid Reasoning Agents (HybridAgents), a new architecture designed to bridge the gap between latent reasoning and text-based interaction. Unlike traditional Large Language Model (LLM) agents that rely on text-based chain-of-thought prompting—which incurs high token costs and latency—HybridAgents perform heavy reasoning within the model’s hidden vector space (latent space) while using natural language only for external actions.

The system utilizes a lightweight HybridLink to transition between text and latent modes, supported by a continuous working memory to maintain context. To optimize this architecture, developers implemented HybridSFT for stabilizing tool-integrated generation and HybridRL to scale reasoning by determining when to use latent space versus text.

Evaluations across 11 benchmarks in fields such as mathematics, science, and medicine indicate that HybridAgent improves performance by 10.4% compared to existing frontier LLM agents. Additionally, the framework offers significant efficiency gains, including a 35.5% to 54.7% reduction in tokens, 1.8x to 2.6x faster inference, and cost savings between 2.7x and 3.7x.