Persuasion Tokens for Editing Factual Knowledge in LLMs

TL;DR

This paper introduces persuasion tokens (P-Tokens), a novel method for efficiently editing factual knowledge in large language models without the need for lengthy demonstrations, showing comparable or better performance than existing techniques.

Contribution

The paper proposes P-Tokens, trained tokens that replicate in-context demonstrations, offering a scalable and effective alternative for knowledge editing in LLMs.

Findings

P-Tokens achieve performance comparable to IKE.

Editing performance is robust to distractors.

Increasing P-Tokens improves editing results.

Abstract

In-context knowledge editing (IKE) is a promising technique for updating Large Language Models (LLMs) with new information. However, IKE relies on lengthy, fact-specific demonstrations which are costly to create and consume significant context window space. In this paper, we introduce persuasion tokens (P-Tokens) -- special tokens trained to replicate the effect of IKE demonstrations, enabling efficient knowledge editing without requiring fact-specific demonstrations. We evaluate P-Tokens across two editing datasets and three LLMs, demonstrating performance comparable to, and often exceeding, IKE. We further find that editing performance is robust to distractors with small negative effects to neighboring facts, and that increasing the number of P-Tokens improves performance. Our work addresses key limitations of IKE and provides a more practical and scalable alternative for editing LLMs.