Adaptive Token Biaser: Knowledge Editing via Biasing Key Entities

TL;DR

The paper introduces ATBias, a decoding technique that improves knowledge editing in large language models by biasing tokens related to key entities, significantly enhancing performance with minimal latency increase.

Contribution

ATBias is a novel decoding method that selectively biases tokens related to key entities, improving in-context editing efficiency and effectiveness in large language models.

Findings

Achieves up to 32.3% improvement over state-of-the-art ICE methods.

Reduces latency by half compared to existing techniques.

Widely applicable to various LLMs with negligible additional cost.

Abstract

The parametric knowledge memorized by large language models (LLMs) becomes outdated quickly. In-context editing (ICE) is currently the most effective method for updating the knowledge of LLMs. Recent advancements involve enhancing ICE by modifying the decoding strategy, obviating the need for altering internal model structures or adjusting external prompts. However, this enhancement operates across the entire sequence generation, encompassing a plethora of non-critical tokens. In this work, we introduce $\textbf{A}$daptive $\textbf{T}$oken $\textbf{Bias}$er ($\textbf{ATBias}$), a new decoding technique designed to enhance ICE. It focuses on the tokens that are mostly related to knowledge during decoding, biasing their logits by matching key entities related to new and parametric knowledge. 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.