Is Model Editing Built on Sand? Revealing Its Illusory Success and Fragile Foundation

TL;DR

The paper challenges the perceived reliability of contemporary model editing by showing that high edit success often rests on semantically shallow shortcuts rather than genuine knowledge updates. It introduces negation- and fact-checking-style evaluations to probe semantic grounding, demonstrating that state-of-the-art edits fail under negation and exhibit large gaps between efficacy and true semantic understanding. By conducting large-scale experiments across multiple models and datasets, the authors reveal pervasive illusory success and advocate for evaluation frameworks that penalize shortcut-driven edits. The work highlights an urgent need to rethink the foundational approach to model editing and to prioritize robust semantic integration over superficial alignment.

Abstract

Large language models (LLMs) inevitably encode outdated or incorrect knowledge. Updating, deleting, and forgetting such knowledge is important for alignment, safety, and other issues. To address this issue, model editing has emerged as a promising paradigm: by precisely editing a small subset of parameters such that a specific fact is updated while preserving other knowledge. Despite its great success reported in previous papers, we find the apparent reliability of editing rests on a fragile foundation and the current literature is largely driven by illusory success. The fundamental goal of steering the model's output toward a target with minimal modification would encourage exploiting hidden shortcuts, rather than utilizing real semantics. This problem directly challenges the feasibility of the current model editing literature at its very foundation, as shortcuts are inherently at odds with robust knowledge integration. Coincidentally, this issue has long been obscured by evaluation frameworks that lack the design of negative examples. To uncover it, we systematically develop a suite of new evaluation methods. Strikingly, we find that state-of-the-art approaches collapse even under the simplest negation queries. Our empirical evidence shows that editing is likely to be based on shortcuts rather than full semantics, calling for an urgent reconsideration of the very basis of model editing before further advancements can be meaningfully pursued.

Figures (5)

• Figure 1: (a) An example of the goal of LLM editing: updating the outdated knowledge with modifying only a small set of parameters (e.g., the blue part). (b) A toy example about current paradigm of model editing is not done on the real semantics.
• Figure 2: A data example from Counterfact.
• Figure 3: A qualitative illustrative example of the experimental failure case under negation (old model knowledge is French, see Figure \ref{['fig: ddenglish']}). Either the model is edited with "XX is YY” or "XX is not YY", and either the test query is "XX is” or "XX is not”, results consistently tend to be "YY".
• Figure 4: The example of steering the output to "gibbon" with shortcuts.
• Figure 5: Examples of prompts used for fact-checking evaluation.