HybridAgent: The AI Agent That Thinks Without Words and Acts Through Text
HybridAgent is a new LLM agent from researchers at Stanford University, Together AI, UIUC and Meta that does most of its thinking inside the model’s hidden space and only switches to text when it needs to call a tool, give an answer or show its work. The paper appeared on arXiv in October 2026. According to the authors’ reported figures, HybridAgent uses 35.5% to 54.7% fewer tokens, runs 1.8x to 2.6x faster and costs 2.7x to 3.7x less than the agents it was compared with, while scoring 10.4% higher on average.

Key Takeaways
- Who: Jiaru Zou, James Zou and colleagues from Stanford, Together AI, UIUC and Meta.
- When: Paper published on arXiv (2610.18861) in October 2026.
- What: The agent switches between “latent mode” (silent thinking in hidden vectors) and “text mode” (normal words and tool calls) in one generation.
- Efficiency: 35.5–54.7% fewer tokens, 1.8–2.6x faster inference, 2.7–3.7x lower cost, according to the paper.
- Accuracy: 10.4% average improvement across 11 benchmarks in math, science, medicine, search and agentic tasks.
- Access: Code is on GitHub and 4B, 8B and 14B checkpoints are on Hugging Face. It is a research release, not a consumer app.
Watch: our 30-second video summary (in Vietnamese).
What Is HybridAgent?
Most AI agents today think the same way they talk: in text. Every reasoning step is written out token by token (a token is a small chunk of a word) before the agent can act. That is easy to read, but slow and expensive.
A separate line of research, called latent reasoning, lets a model think directly in its internal number space, the “hidden states” it computes before choosing a word. That is faster, but latent reasoners had two problems. They could not call tools such as a search engine or Python, and nobody could read what they were thinking.
This new agent tries to get the best of both. The authors describe it as an agent that “thinks in latent space, acts through text.” It thinks silently when that is efficient. It writes text when it must talk to the outside world.
The agent is built on Qwen3-Base models at three sizes: 4 billion, 8 billion and 14 billion parameters. In the experiments it used two tools: a sandboxed Python interpreter and a live search engine.
How Does It Switch Between Latent Thinking and Text?
This is the part the video could not fit. Here is how the switch works, in plain terms.
Text mode
In text mode, the agent behaves like any normal LLM. It picks the next word from its vocabulary, one at a time. It can write a tool call, read the tool’s result and write a final answer.
Entering latent mode
When the model produces a special token, <latent>, it enters latent mode. From that point it stops turning its thoughts into words. Instead, it takes its last-layer hidden state and feeds it straight back in as the next input.
HybridLink: the bridge
Feeding a hidden state back in raw does not work well, because hidden states and word embeddings live in slightly different spaces. So the team added HybridLink, a small adapter that translates the hidden state back into something that looks like a normal input embedding. It is lightweight: roughly 10 million parameters for a 4B model. Because it uses a residual connection, it only has to fix the difference between the two spaces, not learn everything from scratch.
Leaving latent mode
The model keeps thinking silently until its hidden state lines up with an exit marker, </latent>. Then it drops back to text mode. In the experiments each latent span ran for at least 16 and at most 256 steps.
Working memory across turns
Agents work in several turns: think, call a tool, read the result, think again. The agent carries everything (text, latent thoughts and tool results) forward through its own working memory, a persistent cache of the model’s internal state. The authors say this keeps the full context without loss between turns.
Reading the silent thoughts
The most surprising finding: latent thoughts are not a black box. Each latent step can be matched to its nearest word, producing “pseudo-text” that humans can read. The team says these readouts are clear and relevant to the question once the model has gone through its full training. That matters for safety and debugging, because you can audit what the agent was thinking.
How Is It Trained?
Training happens in two stages:
- HybridSFT (supervised fine-tuning) teaches the base model to work together with HybridLink, using 54,000 curated hybrid examples. A step called self-conditioned training lets the model learn from latent thoughts produced by its own HybridLink.
- HybridRL (reinforcement learning) teaches the agent when to think silently and when to act through text, using 12,000 deep reasoning examples and 1,000 agentic search examples.
During reinforcement learning, the 14B model gradually moved more of its thinking into latent mode while getting better at using tools.
HybridAgent Benchmarks and Efficiency Data
The agent was tested on 11 benchmarks:
| Area | Benchmarks |
|---|---|
| Math | AIME 25, AIME 26, MATH500 |
| Knowledge and science | GPQA-Diamond, MedQA, MuSiQue |
| Agentic reasoning | GAIA, Humanity’s Last Exam, WebWalker, SealQA, DeepSearchQA |
It was compared against text-based agent training (SFT, GRPO, ARPO), latent reasoning methods (SoftThinking, Coconut, SoftCoT), frontier agents (GPT-5.6-Luna, Kimi-K3, GLM-5.3-Flash, DeepSeek-V4.1-Flash, QwQ-32B) and agent frameworks (ReAct, Search-o1, WebThinker).
Here are the headline numbers, according to the authors’ reported figures:
| Metric | Reported result |
|---|---|
| Average performance gain | +10.4% |
| Gain over strongest agentic baseline (14B, agentic tasks) | +17.4% average |
| Gain over text/latent baselines at SFT stage | +18.5% |
| Token reduction | 35.5% to 54.7% |
| Inference speed | 1.8x to 2.6x faster |
| Compute cost | 2.7x to 3.7x lower |
| Text tokens vs text-based baselines | 2.7x fewer |
The team also checked that the silent thoughts actually matter. When they replaced latent thoughts with thoughts from another problem or random tokens on MATH500, accuracy fell by roughly 45% to 51%. In other words, the agent really relies on what it thinks in latent mode.
How to Access HybridAgent
The project is an open research release:
- Code: the GitHub repository
togethercomputer/HybridAgentcontains the hybrid generation engine, Python and search tools, both training stages, batch inference and a latent-mode inspector that renders thoughts as pseudo-text. - Models: checkpoints are in the HybridAgent collection on Hugging Face. Each model folder holds the backbone plus HybridLink weights.
- Hardware: the repository recommends 8 or 16 GPUs for HybridSFT and 32 GPUs for HybridRL. The paper’s experiments ran on H100 GPUs.
There is no chat app or hosted product. This is for researchers and developers who are comfortable running models themselves. Standard serving tools like vLLM do not support this kind of hybrid generation out of the box, so the team extended their own infrastructure.
If you want to build agents with tools you can use today, see our guides on Gemini Agent and Claude Code subagents. For running open models locally, our Ollama prompt templates are a good start.
Is Latent Reasoning a Big Deal?
Strengths
- Cheaper and faster agents without giving up tool use.
- Latent thoughts can be read as pseudo-text, which helps auditing.
- Code and models are open, so others can test the claims.
Limitations
- All numbers come from the authors’ own paper. Independent replication is still needed.
- Tested models are 4B to 14B. We do not yet know how this scales to the largest models.
- Training needs a multi-GPU cluster, and normal serving tools do not support it yet.
- Pseudo-text is a nearest-word approximation, not a full transcript of the model’s thinking.
The big idea is simple: an agent does not need to “speak” every thought. If this holds up, future agents could be cheaper to run while staying inspectable.
HybridAgent FAQ
What is HybridAgent?
HybridAgent is an LLM agent that thinks in the model’s hidden (latent) space and switches to text for tool calls and answers. It was built by researchers from Stanford, Together AI, UIUC and Meta.
How much cheaper is HybridAgent?
According to the paper, it lowers compute cost by 2.7x to 3.7x and uses 35.5% to 54.7% fewer tokens than the baselines it was tested against.
Can I read what HybridAgent is thinking?
Yes, approximately. Each of its latent thoughts can be rendered as readable pseudo-text by matching it to the nearest word in the vocabulary.
Is HybridAgent open source?
The code is on GitHub and 4B, 8B and 14B checkpoints are on Hugging Face.
What is HybridLink?
HybridLink is a small adapter, about 10 million parameters for a 4B model, that turns the model’s hidden state back into an input it can keep thinking with.
Sources
- Hybrid Reasoning Agents: LLM Agents that Think in Latent Space, Act through Text (arXiv 2610.18861)
- HybridAgent on GitHub (togethercomputer)
- HybridAgent collection on Hugging Face
Last updated: October 23, 2026