Hybrid Reasoning Agents
HybridAgent moves between latent “thinking” and text/tool use inside one generation process.
The paper proposes a lightweight HybridLink for switching modes and a continuous working memory meant to preserve hybrid context across turns. Its post-training setup uses HybridSFT to stabilize tool-integrated generation, then HybridRL to learn when to reason latently and when to act in text. Across 11 benchmarks at three model scales, the authors report a 10.4% performance gain, 35.5%–54.7% fewer tokens, 1.8x–2.6x faster inference, and 2.7x–3.7x cost savings. Code and data are listed as available. ArXiv · AI/CL/LG's note
The paper proposes a lightweight HybridLink for switching modes and a continuous working memory meant to preserve hybrid context across turns. Its post-training setup uses HybridSFT to stabilize tool-integrated generation, then HybridRL to learn when to reason latently and when to act in text. Across 11 benchmarks at three model scales, the authors report a 10.4% performance gain, 35.5%–54.7% fewer tokens, 1.8x–2.6x faster inference, and 2.7x–3.7x cost savings. Code and data are listed as available. ArXiv · AI/CL/LG's note
score 5