2025/03/11 by Bilgehan Sel, Sel, Bilgehan, Dingcheng Li +9 · 1 citation
Medicine · #Artificial Intelligence (cs.AI) #Automotive and Human Injury Biomechanics #Computation and Language (cs.CL) #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.2503.08919
openalex publication_date 2025/03/11 · openalex created_date 2025/10/13 · openalex updated_date 2026/07/28
Large language models (LLMs) have demonstrated remarkable capabilities across various tasks, but ensuring their safety and alignment with human values remains crucial. Current safety alignment methods, such as supervised fine-tuning and reinforcement learning-based approaches, can exhibit vulnerabilities to adversarial attacks and often result in shallow safety alignment, primarily focusing on preventing harmful content in the initial tokens of the generated output. While methods like resetting can help recover from unsafe generations by discarding previous tokens and restarting the generation process, they are not well-suited for addressing nuanced safety violations like toxicity that may arise within otherwise benign and lengthy generations. In this paper, we propose a novel backtracking method designed to address these limitations. Our method allows the model to revert to a safer generation state, not necessarily at the beginning, when safety violations occur during generation. This approach enables targeted correction of problematic segments without discarding the entire generated text, thereby preserving efficiency. We demonstrate that our method dramatically reduces toxicity appearing through the generation process with minimal impact to efficiency.