A new AI agent named HybridAgent has demonstrated the ability to perform complex, multi-step tasks by toggling between internal, non-textual reasoning and explicit text output. When asked about the corporate headquarters of software engineer Sachin Warrier, the agent spent 30 steps drafting a plan, followed by 56 steps of silent processing before querying an external search engine. After receiving the results, it underwent another 49 steps of silent deliberation to finally output the answer: Mumbai. This research, led by James Zou at Stanford University alongside teams from Together AI and the University of Illinois, was published on arXiv on October 13, 2026, with the source code and model weights made publicly available.

Key Insights
  • Why current reasoning models output every step of their thought process and the financial implications of this approach.
  • The limitations of previous silent-reasoning models in interacting with web-based search engines.
  • The impact on mathematical accuracy when the silent deliberation process of an AI is interrupted.
  • Why safety researchers are concerned about the implications of opaque, unreadable machine thoughts.

What is the financial and technical cost of chain-of-thought processing?

Modern reasoning models generate a lengthy internal monologue before arriving at a final answer, which is composed of units called tokens. Because the model must generate each token sequentially, this process is computationally expensive and slow. While OpenAI does not expose these "thought tokens" to users, they are included in the billable output, often ranging from hundreds to tens of thousands of tokens per query. As this internal chain of thought extends, both the latency and the financial cost for the user increase accordingly.

The concept of silent reasoning is not entirely new. In December 2024, Shibo Hao and his team at Meta introduced a method called Coconut, which passed the internal state—a vector of numbers—directly to the next step without converting it into text. However, this "latent space" reasoning remained isolated from the outside world. Since search engines and Python interpreters require text to function, the 14-billion-parameter Coconut model struggled in real-world scenarios, scoring only 6.8 percent on the GAIA benchmark for multi-step tasks.

How does the HybridLink bridge connect different reasoning modes?

The Stanford research team introduced a compact layer named HybridLink, which serves as a bridge between the model's internal state and the language space. In a 4-billion-parameter model, this bridge accounts for approximately 10 million parameters, or 0.25 percent of the total model size. When the model triggers a specific signal, it enters a silent mode, returning to text only when its internal state is most closely aligned with an output token. This ensures that when the agent needs to perform a search, execute code, or provide a final answer, it can seamlessly convert its silent deliberations into the text required by external tools. All previous steps, including silent deliberations and search results, are maintained in a shared working memory.

The training process involved two phases. First, the model was shown 54,000 example flows to identify which steps should be silent versus written. Subsequently, through reinforcement learning, the model autonomously learned when to pause and when to output text based on success-driven rewards. Although researchers did not explicitly reward the use of silent modes, the 14-billion-parameter model increasingly shifted its reasoning into silence as training progressed, utilizing external tools more effectively. Testing was conducted using the 4B, 8B, and 14B versions of Alibaba's open-source Qwen3 models.

Notes and Sources

  1. Zou, J., Qi, Y., Chen, S., Bianchi, F., Zhu, S., Lu, P., He, J. and Zou, J. (2026). "Hybrid Reasoning Agents". arXiv:2610.18861 (preprint). doi:10.48550/arXiv.2610.18861
  2. Together AI (2026). HybridAgent codebase and HybridAgent-8B-Agentic model. github.com/togethercomputer/HybridAgent
  3. OpenAI (2026). "Reasoning models", developer documentation. developers.openai.com/api/docs/guides/reasoning
  4. Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J. and Tian, Y. (2024). "Training Large Language Models to Reason in a Continuous Latent Space". arXiv:2412.06769. arxiv.org/abs/2412.06769
  5. Korbak, T., Balesni, M., Barnes, E., Bengio, Y. et al. (2025). "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety". arXiv:2507.11473. arxiv.org/abs/2507.11473