AI glossary · Basic · Agents · Language models · Efficiency & infra
Hybrid Reasoning Agents
Hybrid Reasoning Agents (HybridAgents) is an AI agent architecture that reasons in the model's internal numerical representations (hidden representations) instead of writing its thoughts out, and switches to text only when calling tools or giving the final answer. It was proposed in a paper of the same name released on arXiv in October 2026.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents, Language models, Efficiency & infra
- Added
- 2026-10-17
What is different
Today's LLM agents usually reason by writing out everything, including the thinking process (CoT), and call tools through that text as well. This lets people read the process, but it uses many tokens, which makes it slow and expensive. Latent reasoning, by contrast, thinks using only the model's internal representations without writing text, which is efficient, but it has limits: it is hard to use tools directly, and hard to see what the model was thinking. Hybrid reasoning agents try to combine the strengths of both by switching between the two modes as needed within a single generation process.
Three ways of reasoning
Text-based agentsThink and act in text
- Process is easy for people to read
- Tool calls come naturally
- Many tokens: slow and expensive
Latent reasoningThinks in internal representations
- Efficient
- Hard to use tools directly
- Hard to see its thinking
Hybrid reasoning agentsSwitches as needed
- Thinking happens internally
- Tool calls and final answers in text
- Learns when to switch through training
How it works
According to the paper, a lightweight connector called HybridLink handles switching between text mode and latent mode. The agent thinks quietly in latent mode and moves to text mode when it needs to call a tool such as search or give the final answer. The paper says multi-turn conversations and tool results are held in a continuous working memory, so the agent keeps going without losing a context that mixes text and internal representations. The researchers say they designed it so that thoughts in the latent state can also be inspected directly.
How it is trained
Post-training happens in two stages. First, HybridSFT teaches the base model to work together with HybridLink and reliably perform hybrid generation that mixes in tool use. Then reinforcement learning called HybridRL teaches the model to decide for itself when to think internally and when to act in text. In other words, the ability to choose 'how to think' is itself made a learning target.
Why it is getting attention
Agents go through many steps and use many tokens, so speed and cost are a major burden in real services. The researchers report that, evaluated across 3 model sizes and 11 benchmarks spanning math, science, medicine, search and agentic reasoning, performance was on average 10.4% higher than existing agents. In the same evaluation, they claim token usage fell by 35.5–54.7%, inference became 1.8–2.6× faster, and cost was reduced by 2.7–3.7×. They say they have also released the code and data.
Results reported by the researchers
Things to know
This is paper-stage research posted on arXiv on October 13, 2026, and all the numbers above are results reported by the researchers. Reproduction by other researchers or use in real products has not yet been confirmed. The participating institutions (Stanford·Together AI·UIUC·Meta) come from a community post. It also needs separate examination whether 'latent thoughts can be inspected' means they can be understood as easily as a person reading a written CoT. As more of the thinking happens internally, it may become harder to monitor why an agent took a given action.
Example
This site posted, as a news item, a community report that 'researchers from Stanford·Together AI·UIUC·Meta released the Hybrid Reasoning Agents paper, which reasons inside hidden representations and uses text only for tool calls and final answers.' For example, given a question that requires several web searches, an existing agent writes out lengthy thoughts at every step before writing a search query. A hybrid reasoning agent handles those thinking steps internally and outputs only the search queries and the final answer as text, saving tokens and time.