HybridAgent toggles latent reasoning with tools, cuts cost 3.7x

An agent that switches mid-generation between text and silent 'latent' reasoning posts a 10.4% average performance gain over frontier LLM agents while using 35.5% to 54.7% fewer tokens, running 1.8x to 2.6x faster, and costing 2.7x to 3.7x less across 11 benchmarks at three model scales, according to a new arXiv preprint titled 'Hybrid Reasoning Agents'.

The pitch is a toggle. The authors describe a 'lightweight HybridLink' that 'enables seamless transitions between text and latent modes', paired with a 'continuous working memory' preserving 'lossless hybrid context across multiple turns'. The agent reasons silently in latent space when internal thinking is cheaper and surfaces into text when it needs to invoke a tool, which standard latent-reasoning setups cannot do natively.

Training is split into two post-training stages. 'HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with HybridLink', the paper says, and 'HybridRL scales reasoning by learning when to think in latent space and when to act through text'. The 11 benchmarks span mathematics, science, medicine, search, and agentic reasoning.

Code and data are posted at github.com/togethercomputer/HybridAgent. The abstract publishes ranges rather than per-benchmark figures, does not name the specific frontier baselines, and does not identify the three model scales it evaluates.