[Submitted on 13 Oct 2026]
Abstract:How can we build an delegate that thinks beyond individual language, yet motionless acts through it in the external environment? Modern LLM agents motionless logic and engage through text, during latent reasoning offers an productive inner average but lacks native tool admission and reasoning transparency. We extend these two directions alongside Hybrid Reasoning Agents (HybridAgents), which dynamically navigate between latent reasoning and text-based communication throughout a unified generation process. Architecturally, a lightweight HybridLink enables seamless transitions between content and latent modes, during the agent's uninterrupted operating recollection preserves lossless hybrid environment throughout multiple turns. These designs maintain tool-integrated agentic capabilities during making latent thoughts immediately inspectable. Building on this architecture, we additional create dedicated post-training algorithms for hybrid reasoning: HybridSFT stabilizes tool-integrated hybrid generation by instruction the backbone to coordinate alongside HybridLink, afterward HybridRL scales reasoning by learning whenever to think in latent area and whenever to act through text. We measure HybridAgent at 3 example scales throughout 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared alongside frontier LLM Agents and agentic frameworks spanning the two text-based and latent methods, HybridAgent consistently improves achievement by 10.4%, during delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x disbursal savings. Code and data are accessible at this https URL.
Submission history
From: Jiaru Zou [view email]
[v1] Tue, 13 Oct 2026 17:59:31 UTC (8,500 KB)
