Sowing the Wind, Reaping the Whirlwind: The Impact of Editing Language Models

TL;DR

This work investigates knowledge editing in large language models and its safety implications, revealing that injecting accurate information via edits can unintentionally destabilize guardrails. It presents a three-phase methodology—Unethical Q&A generation, edit data construction, and edit-based red teaming—and evaluates safety impacts using two topical settings with $M_{base}$ vs $M_{edited}$ across three datasets, integrating GPT-4 as the evaluator. The results show context- and topic-dependent shifts, with notable increases in unsafe outputs for certain topics (e.g., Advanced Technology to Create Weapons) and when using chain-of-thought prompting, highlighting a nontrivial risk-reward tradeoff in topical red-teaming. The study culminates with the NicheHazardQA benchmark and urges safer editing strategies and stronger guardrails, underscoring the need for further research into ethics-aware model editing and robust safety evaluation.

Abstract

In the rapidly advancing field of artificial intelligence, the concept of Red-Teaming or Jailbreaking large language models (LLMs) has emerged as a crucial area of study. This approach is especially significant in terms of assessing and enhancing the safety and robustness of these models. This paper investigates the intricate consequences of such modifications through model editing, uncovering a complex relationship between enhancing model accuracy and preserving its ethical integrity. Our in-depth analysis reveals a striking paradox: while injecting accurate information is crucial for model reliability, it can paradoxically destabilize the model's foundational framework, resulting in unpredictable and potentially unsafe behaviors. Additionally, we propose a benchmark dataset NicheHazardQA to investigate this unsafe behavior both within the same and cross topical domain. This aspect of our research sheds light on how the edits, impact the model's safety metrics and guardrails. Our findings show that model editing serves as a cost-effective tool for topical red-teaming by methodically applying targeted edits and evaluating the resultant model behavior.

Figures (4)

• Figure 1: Sample output before and after editing.
• Figure 2: Schematic of the data preparation to editing framework.
• Figure 3: Sample responses obtained from $M_{edited}$ and $M_{base}$ models for unethical questions across different datasets.
• Figure 4: This table depicts the improvement in tasks achieved by using ROME to edit the LLaMA-2 (7B) model, highlighting the progressive impact as the number of edits increases.