Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget.

Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget.

Due to the high cost of retraining from scratch, researchers attempt to employ machine unlearning to remove specific content from LLMs while preserving the overall performance. In this paper, we discuss several issues in machine unlearning for LLMs and provide our insights on possible approaches. To address the issue of inadequate evaluation of model outputs after unlearning, we introduce three additional metrics to evaluate token diversity, sentence semantics, and factual correctness. We then categorize unlearning methods into untargeted and targeted, and discuss their issues respectively. Specifically, the behavior that untargeted unlearning attempts to approximate is unpredictable and may involve hallucinations, and existing regularization is insufficient for targeted unlearning. To alleviate these issues, we propose using the objective of maximizing entropy (ME) for untargeted unlearning and incorporate answer preservation (AP) loss as regularization for targeted unlearning. Experimental results across three scenarios, i.e., fictitious unlearning, continual unlearning, and real-world unlearning, demonstrate the effectiveness of our approaches. In recent years, large language models (LLMs) have undergone rapid development, demonstrating impressive capabilities across a wide range of applications, from natural language processing to complex problem-solving. These concerns are particularly relevant within legal and regulatory frameworks, such as the Right to be Forgotten (Dang, 2021), which aims to empower individuals to have unauthorized data erased from digital records. Addressing these issues is crucial for ensuring the responsible deployment of LLMs in real-world applications. Due to the high cost of retraining LLMs, researchers have explored machine unlearning techniques, namely LLM unlearning (Cao & Yang, 2015; Bourtoule et al., 2021; Yao et al., 2023). The typical paradigm involves fine-tuning the target LLM on a specified set, known as the forget set, to obtain an unlearned model. As described in (Maini et al., 2024; Jin et al., 2024), the unlearned model should meet two primary goals: 1) it should not reveal any information contained in the forget set, and 2) it should maintain performance on the neighbor set, which has a distribution similar to the forget set but is not the target of unlearning, as well as on other tasks with general knowledge. While the first goal is generally easier to achieve, the main challenge lies in meeting the second goal (Liu et al., 2024b; Maini et al., 2024; Zhang et al., 2024a; Ji et al., 2024; Shi et al., 2024a; Wang et al., 2024c). In this paper, we have a closer look at machine unlearning for LLMs. We note that most prior studies (Maini et al., 2024; Ji et al., 2024; Jia et al., 2024; Jin et al., 2024; Shi et al., 2024a) primarily rely on ROUGE (Lin, 2004) as the sole metric for evaluating the output of unlearned models.

Formal logic has long been applied to natural language reasoning, but this approach can sometimes lead to conclusions that, while logically entailed, are factually inconsistent with the premises or are not typically inferred by humans. This study introduces the concept of "rulebreakers", which refers to instances where logical entailment diverges from factually acceptable inference. We present RULEBREAKERS, a novel dataset for evaluating Large Language Models' (LLMs) ability to distinguish between rulebreakers and non-rulebreakers. Focusing on modus tollens and disjunctive syllogism, we assess six state-of-the-art LLMs using RULEBREAKERS, measuring their performance in terms of token-level exact accuracy and model confidence. Our findings reveal that while most models perform poorly to moderately in recognizing rulebreakers, they demonstrate a latent ability to distinguish rulebreakers when assessed by their confidence levels. Further analysis suggests that the failure to recognize rulebreakers is potentially associated with the models' world knowledge and their attention distribution patterns. This research highlights the limitation of LLMs' reasoning capabilities, and contributes to the ongoing discussion on reasoning in LLMs.

We present a principled approach to provide LLM-based evaluation with a rigorous guarantee of human agreement. We first propose that a reliable evaluation method should not uncritically rely on model preferences for pairwise evaluation, but rather assess the confidence of judge models and selectively decide when to trust its judgement. We then show that under this selective evaluation framework, human agreement can be provably guaranteed -- such that the model evaluation aligns with that of humans to a user-specified agreement level. As part of our framework, we also introduce Simulated Annotators, a novel confidence estimation method that significantly improves judge calibration and thus enables high coverage of evaluated instances. Finally, we propose Cascaded Selective Evaluation, where we use cheaper models as initial judges and escalate to stronger models only when necessary -- again, while still providing a provable guarantee of human agreement. Experimental results show that Cascaded Selective Evaluation guarantees strong alignment with humans, far beyond what LLM judges could achieve without selective evaluation. For example, on a subset of Chatbot Arena where GPT-4 almost never achieves 80% human agreement, our method, even while employing substantially cost-effective models such as Mistral-7B, guarantees over 80% human agreement with almost 80% test coverage.

The compelling goal of eradicating undesirable data behaviors, while preserving usual model functioning, underscores the significance of machine unlearning within the domain of large language models (LLMs). Recent research has begun to approach LLM unlearning via gradient ascent (GA) -- increasing the prediction risk for those training strings targeted to be unlearned, thereby erasing their parameterized responses. Despite their simplicity and efficiency, we suggest that GA-based methods face the propensity towards excessive unlearning, resulting in various undesirable model behaviors, such as catastrophic forgetting, that diminish their practical utility. In this paper, we suggest a set of metrics that can capture multiple facets of real-world utility and propose several controlling methods that can regulate the extent of excessive unlearning. Accordingly, we suggest a general framework to better reflect the practical efficacy of various unlearning methods -- we begin by controlling the unlearning procedures/unlearned models such that no excessive unlearning occurs and follow by the evaluation for unlearning efficacy. Our experimental analysis on established benchmarks revealed that GA-based methods are far from perfect in practice, as strong unlearning is at the high cost of hindering the model utility. We conclude that there is still a long way towards practical and effective LLM unlearning, and more efforts are required in this field.

Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a large language model should not know is important for ensuring alignment and thus safe use. However, accurately and efficiently unlearning knowledge from an LLM remains challenging due to the potential collateral damage caused by the fuzzy boundary between retention and forgetting, and the large computational requirements for optimization across state-of-the-art models with hundreds of billions of parameters. In this work, we present Embedding-COrrupted (ECO) Prompts, a lightweight unlearning framework for large language models to address both the challenges of knowledge entanglement and unlearning efficiency. Instead of relying on the LLM itself to unlearn, we enforce an unlearned state during inference by employing a prompt classifier to identify and safeguard prompts to forget. We learn corruptions added to prompt embeddings via zeroth order optimization toward the unlearning objective offline and corrupt prompts flagged by the classifier during inference. We find that these embedding-corrupted prompts not only lead to desirable outputs that satisfy the unlearning objective but also closely approximate the output from a model that has never been trained on the data intended for forgetting. Through extensive experiments on unlearning, we demonstrate the superiority of our method in achieving promising unlearning at nearly zero side effects in general domains and domains closely related to the unlearned ones. Additionally, we highlight the scalability of our method to 100 LLMs, ranging from 0.5B to 236B parameters, incurring no additional cost as the number of parameters increases.

Pretrained large language models have revolutionized many applications but still face challenges related to cultural bias and a lack of cultural commonsense knowledge crucial for guiding cross-culture communication and interactions. Recognizing the shortcomings of existing methods in capturing the diverse and rich cultures across the world, this paper introduces a novel approach for massively multicultural knowledge acquisition. Specifically, our method strategically navigates from densely informative Wikipedia documents on cultural topics to an extensive network of linked pages. Leveraging this valuable source of data collection, we construct the CultureAtlas dataset, which covers a wide range of sub-country level geographical regions and ethnolinguistic groups, with data cleaning and preprocessing to ensure textual assertion sentence self-containment, as well as fine-grained cultural profile information extraction. Our dataset not only facilitates the evaluation of language model performance in culturally diverse contexts but also serves as a foundational tool for the development of culturally sensitive and aware language models. Our work marks an important step towards deeper understanding and bridging the gaps of cultural disparities in AI, to promote a more inclusive and balanced representation of global cultures in the digital domain.

Multi-modal entity alignment (MMEA) aims to discover identical entities across different knowledge graphs (KGs) whose entities are associated with relevant images. However, current MMEA algorithms rely on KG-level modality fusion strategies for multi-modal entity representation, which ignores the variations of modality preferences of different entities, thus compromising robustness against noise in modalities such as blurry images and relations. This paper introduces MEAformer, a multi-modal entity alignment transformer approach for meta modality hybrid, which dynamically predicts the mutual correlation coefficients among modalities for more fine-grained entity-level modality fusion and alignment. Experimental results demonstrate that our model not only achieves SOTA performance in multiple training scenarios, including supervised, unsupervised, iterative, and low-resource settings, but also has a limited number of parameters, efficient runtime, and interpretability. Our code is available at https://github.com/zjukg/MEAformer.

Activation functions play a significant role in neural network design by enabling non-linearity. The choice of activation function was previously shown to influence the properties of the resulting loss landscape. Understanding the relationship between activation functions and loss landscape properties is important for neural architecture and training algorithm design. This study empirically investigates neural network loss landscapes associated with hyperbolic tangent, rectified linear unit, and exponential linear unit activation functions. Rectified linear unit is shown to yield the most convex loss landscape, and exponential linear unit is shown to yield the least flat loss landscape, and to exhibit superior generalisation performance. The presence of wide and narrow valleys in the loss landscape is established for all activation functions, and the narrow valleys are shown to correlate with saturated neurons and implicitly regularised network configurations.