ZihaoZhu/BoT-QwQ-32B

To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models

This model, BoT-QwQ-32B, is a 32.8 billion parameter model developed by Zihao Zhu and his team, stemming from their research paper "To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models." The core focus is on identifying and exploiting a critical flaw in Large Reasoning Models (LRMs) where their 'thinking' or reasoning process can be bypassed by manipulating specific delimiter tokens. This phenomenon is termed Unthinking Vulnerability.

Key Capabilities & Research Areas

• Breaking of Thought (BoT): Investigates how to maliciously bypass LRM reasoning. This includes:
  • Training-based BoT: Injecting backdoors during fine-tuning (SFT and DPO) to force reasoning bypass.
  • Training-free BoT: Exploiting the vulnerability during inference using adversarial attacks (single, universal, and transfer attacks).
• Monitoring of Thought (MoT): Explores beneficial applications of the Unthinking Vulnerability to enhance LRM efficiency and safety alignment.
  • Efficiency Enhancement: Addressing overthinking by monitoring the thinking process.
  • Safety Enhancement: Improving safety alignment through similar monitoring mechanisms.

Good for

• Academic Research: Ideal for researchers studying LRM security, robustness, and the fundamental mechanisms of reasoning in large language models.
• Vulnerability Analysis: Useful for understanding and demonstrating potential attack vectors against reasoning capabilities in LLMs.
• Developing Robust LLMs: Provides a framework and insights for building more resilient and secure reasoning systems.
• Exploring LLM Efficiency: Offers methods to potentially optimize LRM performance by managing or bypassing unnecessary reasoning steps.