The Performance of "Thought": An Exploratory Study on LLM Visualized Reasoning

Abstract: Is the "thinking process" of modern LLMs a faithful log of reasoning, or a manipulable performance? This study introduces "Protocol Injection" to probe the authenticity of visualized reasoning in DeepSeek, Gemini, Claude, and ChatGPT. We uncover the "Safety Arbitration Chamber" mechanism and the risks of "False Transparency".

📥 Download Paper

📄 Read the Full Paper (PDF)

---

🚀 Key Findings (Highlights)

Our research (N=50 automated probes) reveals a "Functional Spectrum" of visualized reasoning:

1. Decoupling: The "Thought" and "Answer" are logically separable.
2. Safety Arbitration Chamber: In models like Claude, the thinking block is hijacked by safety layers for real-time censorship.
3. False Transparency: We demonstrate a novel attack vector where "trusted" thinking processes can be manipulated to carry hidden payloads.

🛠️ Methodology: Protocol Injection

We define "Protocol Injection" not as a jailbreak, but as a diagnostic probe. By injecting conflicting meta-instructions, we force the model to reveal its internal hierarchy between "Performance" and "Safety".

(Code and automated testing scripts coming soon)

📖 Citation

If you find this research useful, please cite it as:

@misc{wu2025performance,
  title={The Performance of "Thought": An Exploratory Study on the Authenticity, Manipulability, and "False Transparency" Risks of LLM Visualized Reasoning},
  author={Wu, Yihu},
  year={2025},
  howpublished={GitHub Repository},
  url={[https://github.com/YOUR_GITHUB_USERNAME/LLM-Thinking-Protocol-Injection](https://github.com/YOUR_GITHUB_USERNAME/LLM-Thinking-Protocol-Injection)}
}

📬 Contact

TyloAI Research [https://www.tyloai.com]