The Fall of ROME: Understanding the Collapse of LLMs in Model Editing

Despite significant progress in model editing methods, their application in real-world scenarios remains challenging as they often cause large language models (LLMs) to collapse. Among them, ROME is particularly concerning, as it could disrupt LLMs with only a single edit. In this paper, we study the root causes of such collapse. Through extensive analysis, we identify two primary factors that contribute to the collapse: i) inconsistent handling of prefixed and unprefixed keys in the parameter update equation may result in very small denominators, causing excessively large parameter updates; ii) the subject of collapse cases is usually the first token, whose unprefixed key distribution significantly differs from the prefixed key distribution in autoregressive transformers, causing the aforementioned issue to materialize. To validate our analysis, we propose a simple yet effective approach: uniformly using prefixed keys during editing phase and adding prefixes during the testing phase. The experimental results show that the proposed solution can prevent model collapse while maintaining the effectiveness of the edits.

Unlink to Unlearn: Simplifying Edge Unlearning in GNNs

As concerns over data privacy intensify, unlearning in Graph Neural Networks (GNNs) has emerged as a prominent research frontier in academia. This concept is pivotal in enforcing the right to be forgotten, which entails the selective removal of specific data from trained GNNs upon user request. Our research focuses on edge unlearning, a process of particular relevance to real-world applications, owing to its widespread applicability. Current state-of-the-art approaches like GNNDelete can eliminate the influence of specific edges, yet our research has revealed a critical limitation in these approaches, termed over-forgetting. It occurs when the unlearning process inadvertently removes excessive information beyond specific data, leading to a significant decline in prediction accuracy for the remaining edges. To address this issue, we have identified the loss functions of GNNDelete as the primary source of the over-forgetting phenomenon. Furthermore, our analysis also suggests that loss functions may not be essential for effective edge unlearning. Building on these insights, we have simplified GNNDelete to develop Unlink-to-Unlearn (UtU), a novel method that facilitates unlearning exclusively through unlinking the forget edges from graph structure. Our extensive experiments demonstrate that UtU delivers privacy protection on par with that of a retrained model while preserving high accuracy in downstream tasks. Specifically, UtU upholds over 97.3% of the retrained model's privacy protection capabilities and 99.8% of its link prediction accuracy. Meanwhile, UtU requires only constant computational demands, underscoring its advantage as a highly lightweight and practical edge unlearning solution.

Prob2Vec: Mathematical Semantic Embedding for Problem Retrieval in Adaptive Tutoring

We propose a new application of embedding techniques for problem retrieval in adaptive tutoring. The objective is to retrieve problems whose mathematical concepts are similar. There are two challenges: First, like sentences, problems helpful to tutoring are never exactly the same in terms of the underlying concepts. Instead, good problems mix concepts in innovative ways, while still displaying continuity in their relationships. Second, it is difficult for humans to determine a similarity score that is consistent across a large enough training set. We propose a hierarchical problem embedding algorithm, called Prob2Vec, that consists of abstraction and embedding steps. Prob2Vec achieves 96.88\% accuracy on a problem similarity test, in contrast to 75\% from directly applying state-of-the-art sentence embedding methods. It is interesting that Prob2Vec is able to distinguish very fine-grained differences among problems, an ability humans need time and effort to acquire. In addition, the sub-problem of concept labeling with imbalanced training data set is interesting in its own right. It is a multi-label problem suffering from dimensionality explosion, which we propose ways to ameliorate. We propose the novel negative pre-training algorithm that dramatically reduces false negative and positive ratios for classification, using an imbalanced training data set.