Dynamic Evil Score-Guided Decoding: An Efficient Decoding Framework For Red-Team Model

Large language models (LLMs) have achieved significant advances, but can potentially generate harmful content such as social biases, extremism, and misinformation. Red teaming is a promising approach to enhance model safety by generating adversarial prompts to test and improve model robustness. However, existing red-teaming methods often require expensive fine-tuning, especially for large LLMs. We propose the Dynamic Evil Score-Guided Decoding framework (DESGD), an efficient red-teaming method that does not increase computational cost with the target model size. DESGD framework introduces the concept of an 'evil score,' which can dynamically evaluate the potential of tokens to contribute to harmful outputs in decoding phrases. We fine-tune a small model using an adversarial dataset and calculate the evil score based on the difference in logits vector before and after fine-tuning. Then, we adjust the logits of the large target model according to the evil score. The results of the experiment show that our method achieves an ASR of 92. 83% on the Llama-3.2-3B-Instruct model, compared to 83.48% with adversarial fine-tuning while using lower computational resources. Similarly, on the Qwen2.5-3B-Instruct model, DESGD reaches an ASR of 88.62%, outperforming adversarial fine-tuning (77.56%).