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

An ordinary LLM agent writes out every intermediate reasoning step as text before it acts. Latent reasoning skips that: instead of turning its internal state into a word and reading the word back in, the model feeds the internal vector itself into its next step. That saves decoding, but a vector cannot call a tool and a person cannot read it. Hybrid Reasoning Agents, a preprint from researchers at Stanford, UIUC and Together AI, builds an agent that switches between the two inside one generation. The agent writes normal text until it emits a special token that opens a latent span. From then on, a small add-on the authors call HybridLink (about 10M parameters for a 4B model) maps each step's final hidden state back into the space the model uses for input words, so the model keeps thinking without producing tokens. It drops back to text when that mapped vector lands closest to the embedding of the special token that closes the span, and tool calls such as Python or search are always made in text. The same nearest-token lookup lets an observer turn any latent step into readable output the authors call pseudo-text. Because a mix of text and vectors cannot be rebuilt by re-tokenizing, the agent carries the whole history from turn to turn in its attention cache, which the paper calls its working memory.

The agents are post-trained from Qwen3 base models at 4B, 8B and 14B. Supervised fine-tuning uses 54K trajectories converted from text agent data, with some reasoning steps reassigned to latent mode, and a self-conditioned step in which the model learns to reason over the vectors its own HybridLink produces rather than idealized ones. Reinforcement learning needed a workaround: standard methods compare the probability of each generated token, and latent steps have none. The authors show that updating only on the sampled text tokens, while replaying the recorded latent vectors exactly, is enough to optimize the whole trajectory. The reward covers correctness, valid formatting and use of more than one tool, with no reward for choosing either mode. Even so, the 14B agent moved more of its generation into latent mode as training went on while making more tool calls.

Across 11 benchmarks in math, science, medicine, search and agentic tasks, the authors report that under matched model size and training budget, their method beats text and latent baselines by 18.5% at the fine-tuning stage and 10.4% after reinforcement learning, on average. Against larger frontier agents the picture is mixed. The 14B agent averages 52.4 on GAIA against 41.7 for Kimi-K3 and 37.9 for GPT-5.6-Luna, but scores 16.0 on Humanity's Last Exam, where those two reach 31.4 and 29.0. The efficiency gains are clearer: 2.7 times fewer text tokens, up to 54% fewer generation steps and 2.1 times faster inference than text-based baselines. The paper also estimates costs 2.7 to 3.7 times lower, but that estimate assumes $2.50 per H100-hour and leaves out prompt processing, context-dependent attention and tool execution. To check that the latent steps do real work, the authors swapped in latent vectors from a different problem, which cut MATH500 accuracy by 44.9% to 47.0%. Forcing the agent to reason only in text dropped the 14B model from 88.2 to 78.8 and the 8B model from 84.6 to 80.4.

The paper's evidence on whether hiding reasoning in vectors hurts oversight comes from one test. On 560 math problems, an outside observer saw only the text, tool calls and tool outputs, with latent spans hidden. The 8B agent exposed 302 text tokens per problem against 596 for a matched text agent. Monitor models asked to predict from that visible trace whether an answer was right did no worse on AUROC (a score for how well the monitor separates right answers from wrong ones): across the 8B and 14B agents and a weak and a strong monitor, the hybrid agent scored higher in three of the four comparisons and 0.001 lower in the fourth. In the discussion section, the authors argue that because latent thoughts can only be read through a HybridLink the provider holds, a provider could monitor the full reasoning while users see only the text, but that is a proposal, not a tested result. Together AI has also published HybridAgent 8B Agentic on Hugging Face, an MIT-licensed fine-tune of Qwen3-8B-Base.