Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

The team trained 4B, 8B and 14B models, all built on Qwen3-Base models, and tested them on 11 benchmarks. Against text-based and latent baselines, the average gain is 10.4%, with 35.5% to 54.7% fewer tokens, 1.8x to 2.6x faster inference and 2.7x to 3.7x lower cost. A special token opens the silent mode, and a small adapter called HybridLink, roughly 10M parameters for a 4B model, feeds the model's hidden states back in as inputs. Latent reasoning is not new. Latent chain-of-thought, introduced as Coconut, replaces sampled text tokens with continuous last-layer hidden states fed back autoregressively. The authors target its weak points: latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. Early reaction to the paper is cautiously enthusiastic about the token and latency savings, but a recurring worry is that hiding reasoning in latent vectors makes debugging and monitoring agent behavior harder. For a team building an AI company, the code and an 8B agentic checkpoint are public, so the claims can be tested on your own agent workloads. The numbers come from the authors' own evaluation, and the MATH500 comparison is one result at one model scale, so it does not establish the same difference across all 11 tests or model sizes. Also note that loading the backbone with transformers alone gives a text-only model; hybrid generation with latent mode and tools needs the HybridAgent code.