This paper introduces Adaptive Token Biaser (ATBIAS), a new decoding technique for enhancing in-context editing (ICE) in large language models (LLMs). The main goal of ATBIAS is to improve the editing capabilities of ICE by focusing on tokens related to new knowledge and biasing their logits while decreasing those of parametric knowledge. ATBIAS achieves this by first extracting key entities from new and parametric knowledge, then filtering tokens based on their probability and ranking, and finally biasing the logits of tokens related to key entities. The framework of ATBIAS is shown in Figure 2, and it focuses more on the matched tokens rather than the entire generated sequence. This approach reduces computational costs and minimizes the risk of introducing errors. Experimental results show that ATBIAS significantly enhances ICE performance, achieving up to a 32.3% improvement over state-of-the-art ICE methods while incurring only half the latency. ATBIAS not only improves the knowledge editing capabilities of ICE but can also be widely applied to LLMs with negligible cost. The paper also discusses the challenges of editing stubborn knowledge, which is difficult to change due to its strong pre-training confidence. ATBIAS addresses this by focusing on key tokens rather than the entire sequence, leading to more effective editing. The paper also presents ablation studies on various components of ATBIAS, including the N-gram decomposition, probabilistic constraint, ranking constraint, and bias coefficients. The results show that ATBIAS achieves the best performance with specific values of these parameters. The paper concludes that ATBIAS is an effective method for enhancing ICE and has significant potential for real-world applications.This paper introduces Adaptive Token Biaser (ATBIAS), a new decoding technique for enhancing in-context editing (ICE) in large language models (LLMs). The main goal of ATBIAS is to improve the editing capabilities of ICE by focusing on tokens related to new knowledge and biasing their logits while decreasing those of parametric knowledge. ATBIAS achieves this by first extracting key entities from new and parametric knowledge, then filtering tokens based on their probability and ranking, and finally biasing the logits of tokens related to key entities. The framework of ATBIAS is shown in Figure 2, and it focuses more on the matched tokens rather than the entire generated sequence. This approach reduces computational costs and minimizes the risk of introducing errors. Experimental results show that ATBIAS significantly enhances ICE performance, achieving up to a 32.3% improvement over state-of-the-art ICE methods while incurring only half the latency. ATBIAS not only improves the knowledge editing capabilities of ICE but can also be widely applied to LLMs with negligible cost. The paper also discusses the challenges of editing stubborn knowledge, which is difficult to change due to its strong pre-training confidence. ATBIAS addresses this by focusing on key tokens rather than the entire sequence, leading to more effective editing. The paper also presents ablation studies on various components of ATBIAS, including the N-gram decomposition, probabilistic constraint, ranking constraint, and bias coefficients. The results show that ATBIAS achieves the best performance with specific values of these parameters. The paper concludes that ATBIAS is an effective method for enhancing ICE and has significant potential for real-world applications.