Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Persuasion Tokens for Editing Factual Knowledge in LLMs

Paul Youssef, Christin Seifert, Jörg Schlötterer

TL;DR

Persuasion Tokens introduce P-Tokens as a compact alternative to in-context knowledge editing (IKE) for updating factual knowledge in LLMs. By embedding BEGIN_EDIT and END_EDIT tokens and optimizing their embeddings via a KL-divergence objective, P-Tokens replicate the effect of lengthy IKE demonstrations with far shorter prompts. The approach is augmented with training signals from paraphrases, neighboring prompts, distractors, and unrelated prompts, and evaluated on CounterFact and zsRE across multiple LLMs. Results show P-Tokens either match or exceed IKE performance, improve efficiency, and exhibit robustness to distractors, making knowledge editing more practical at scale, albeit with initial training costs and potential misuse considerations.

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.

Persuasion Tokens for Editing Factual Knowledge in LLMs

TL;DR

Persuasion Tokens introduce P-Tokens as a compact alternative to in-context knowledge editing (IKE) for updating factual knowledge in LLMs. By embedding BEGIN_EDIT and END_EDIT tokens and optimizing their embeddings via a KL-divergence objective, P-Tokens replicate the effect of lengthy IKE demonstrations with far shorter prompts. The approach is augmented with training signals from paraphrases, neighboring prompts, distractors, and unrelated prompts, and evaluated on CounterFact and zsRE across multiple LLMs. Results show P-Tokens either match or exceed IKE performance, improve efficiency, and exhibit robustness to distractors, making knowledge editing more practical at scale, albeit with initial training costs and potential misuse considerations.

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.
Paper Structure (31 sections, 1 equation, 3 figures, 8 tables)

This paper contains 31 sections, 1 equation, 3 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Background and Problem Statement
  3. In-context Knowledge Editing.
  4. Problem Statement.
  5. Method
  6. Paraphrases:
  7. Neighbors:
  8. Distractors:
  9. Other:
  10. Experimental Setup
  11. Datasets.
  12. Evaluation.
  13. Baselines.
  14. Number of P-Tokens.
  15. Models.
  16. ...and 16 more sections

Figures (3)

  • Figure 1: In-context knowledge editing (IKE) relies on complex demonstrations and leads to slower inference. Persuasion tokens (P-Tokens) eliminate the need for long demonstrations and lead to faster inference.
  • Figure 2: Performance across different numbers of P-Tokens (Left: CounterFact, Right: zsRE). On CounterFact, Qwen2.5-7B benefits from increasing the number of tokens, especially on PS and NS. On zsRE, Llama3's Efficacy and Paraphrase increase when the number of P-Tokens increases.
  • Figure 3: An example of $p_{{}_{IKE}}(s,r,o')$ that changes the mother tongue of Danielle Darrieux from French to English.