With growing demands for data privacy and model robustness, graph unlearning (GU), which erases the influence of specific data on trained GNN models, has gained significant attention. However, existing exact unlearning methods suffer from either low efficiency or poor model performance. While being more utility-preserving and efficient, current approximate unlearning methods are not applicable in the zero-glance privacy setting, where the deleted samples cannot be accessed during unlearning due to immediate deletion requested by regulations. Besides, these approximate methods, which try to directly perturb model parameters still involve high privacy concerns in practice. To fill the gap, we propose Transferable Condensation Graph Unlearning (TCGU), a data-centric solution to zero-glance graph unlearning. Specifically, we first design a two-level alignment strategy to pre-condense the original graph into a small yet utility-preserving dataset. Upon receiving an unlearning request, we fine-tune the pre-condensed data with a low-rank plugin, to directly align its distribution with the remaining graph, thus efficiently revoking the information of deleted data without accessing them. A novel similarity distribution matching approach and a discrimination regularizer are proposed to effectively transfer condensed data and preserve its utility in GNN training, respectively. Finally, we retrain the GNN on the transferred condensed data. Extensive experiments on 6 benchmark datasets demonstrate that TCGU can achieve superior performance in terms of model utility, unlearning efficiency, and unlearning efficacy than existing GU methods.

Large language models (LLMs), like ChatGPT, have shown that even trained with noisy prior data, they can generalize effectively to new tasks through in-context learning (ICL) and pre-training techniques. Motivated by this, we explore whether a similar approach can be applied to scientific foundation models (SFMs). Our methodology is structured as follows: (i) we collect low-cost physics-informed neural network (PINN)-based approximated prior data in the form of solutions to partial differential equations (PDEs) constructed through an arbitrary linear combination of mathematical dictionaries; (ii) we utilize Transformer architectures with self and cross-attention mechanisms to predict PDE solutions without knowledge of the governing equations in a zero-shot setting; (iii) we provide experimental evidence on the one-dimensional convection-diffusion-reaction equation, which demonstrate that pre-training remains robust even with approximated prior data, with only marginal impacts on test accuracy. Notably, this finding opens the path to pre-training SFMs with realistic, low-cost data instead of (or in conjunction with) numerical high-cost data. These results support the conjecture that SFMs can improve in a manner similar to LLMs, where fully cleaning the vast set of sentences crawled from the Internet is nearly impossible. In developing large-scale models, one fundamental challenge is the inherent noisiness of the data used for training. Whether dealing with natural language, scientific data, or other domains, large datasets almost inevitably contain noise. Large language models (LLMs), such as ChatGPT, present an interesting paradox: despite being trained on noisy datasets, they consistently produce remarkably clean and coherent output.

Fairness in artificial intelligence and machine learning (AI/ML) models is becoming critically important, especially as decisions made by these systems impact diverse groups. In education, a vital sector for all countries, the widespread application of AI/ML systems raises specific concerns regarding fairness. Current research predominantly focuses on fairness for individual sensitive features, which limits the comprehensiveness of fairness assessments. This paper introduces FAIREDU, a novel and effective method designed to improve fairness across multiple sensitive features. Through extensive experiments, we evaluate FAIREDU effectiveness in enhancing fairness without compromising model performance. The results demonstrate that FAIREDU addresses intersectionality across features such as gender, race, age, and other sensitive features, outperforming state-of-the-art methods with minimal effect on model accuracy. The paper also explores potential future research directions to enhance further the method robustness and applicability to various machine-learning models and datasets.

Machine-learned surrogate models to accelerate lengthy computer simulations are becoming increasingly important as engineers look to streamline the product design cycle. In many cases, these approaches offer the ability to predict relevant quantities throughout a geometry, but place constraints on the form of the input data. In a world of diverse data types, a preferred approach would not restrict the input to a particular structure. In this paper, we propose Topology-Agnostic Graph U-Net (TAG U-Net), a graph convolutional network that can be trained to input any mesh or graph structure and output a prediction of a target scalar field at each node. The model constructs coarsened versions of each input graph and performs a set of convolution and pooling operations to predict the node-wise outputs on the original graph. By training on a diverse set of shapes, the model can make strong predictions, even for shapes unlike those seen during training. A 3-D additive manufacturing dataset is presented, containing Laser Powder Bed Fusion simulation results for thousands of parts. The model is demonstrated on this dataset, and it performs well, predicting both 2-D and 3-D scalar fields with a median R-squared > 0.85 on test geometries. Code and datasets are available online.

We introduce a new vulnerability that exploits fixed points in autoregressive models and use it to craft queries that never halt, i.e. an LLM output that does not terminate. More precisely, for what we call non-halting queries, the LLM never samples the end-of-string token (). We rigorously analyze the conditions under which the non-halting anomaly presents itself. In particular, at temperature zero, we prove that if a repeating (cyclic) sequence of tokens is observed at the output beyond the context size, then the LLM does not halt. We demonstrate the non-halting anomaly in a number of experiments performed in base (unaligned) models where repeating tokens immediately lead to a non-halting cyclic behavior as predicted by the analysis. Further, we develop a simple recipe that takes the same fixed points observed in the base model and creates a prompt structure to target aligned models. We study the recipe behavior in bypassing alignment in a number of LLMs including GPT-4o, llama-3-8b-instruct, and gemma-2-9b-it where all models are forced into a non-halting state. Further, we demonstrate the recipe's success in sending most major models released over the past year into a non-halting state with the same simple prompt even at higher temperatures. Further, we study direct inversion based techniques to craft new short prompts to induce the non-halting state. Our experiments with the gradient search based inversion technique ARCA show that non-halting is prevalent across models and may be easily induced with a few input tokens. While its impact on the reliability of hosted systems can be mitigated by configuring a hard maximum token limit in the sampler, the non-halting anomaly still manages to break alignment. This underlines the need for further studies and stronger forms of alignment against non-halting anomalies.

Accommodating human preferences is essential for creating AI agents that deliver personalized and effective interactions. Recent work has shown the potential for LLMs to infer preferences from user interactions, but they often produce broad and generic preferences, failing to capture the unique and individualized nature of human preferences. This paper introduces PREDICT, a method designed to enhance the precision and adaptability of inferring preferences. PREDICT incorporates three key elements: (1) iterative refinement of inferred preferences, (2) decomposition of preferences into constituent components, and (3) validation of preferences across multiple trajectories. We evaluate PREDICT on two distinct environments: a gridworld setting and a new text-domain environment (PLUME).

While LLMs have emerged as performant architectures for reasoning tasks, their compositional generalization capabilities have been questioned. In this work, we introduce a Compositional Generalization Challenge for Graph-based Commonsense Reasoning (CGGC) that goes beyond previous evaluations that are based on sequences or tree structures - and instead involves a reasoning graph: It requires models to generate a natural sentence based on given concepts and a corresponding reasoning graph, where the presented graph involves a previously unseen combination of relation types. To master this challenge, models need to learn how to reason over relation tupels within the graph, and how to compose them when conceptualizing a verbalization. We evaluate seven well-known LLMs using in-context learning and find that performant LLMs still struggle in compositional generalization. We investigate potential causes of this gap by analyzing the structures of reasoning graphs, and find that different structures present varying levels of difficulty for compositional generalization. Arranging the order of demonstrations according to the structures' difficulty shows that organizing samples in an easy-to-hard schema enhances the compositional generalization ability of LLMs.

In reinforcement learning, if the agent's reward differs from the designers' true utility, even only rarely, the state distribution resulting from the agent's policy can be very bad, in theory and in practice. When RL policies would devolve into undesired behavior, a common countermeasure is KL regularization to a trusted policy ("Don't do anything I wouldn't do"). All current cutting-edge language models are RL agents that are KL-regularized to a "base policy" that is purely predictive. Unfortunately, we demonstrate that when this base policy is a Bayesian predictive model of a trusted policy, the KL constraint is no longer reliable for controlling the behavior of an advanced RL agent. We demonstrate this theoretically using algorithmic information theory, and while systems today are too weak to exhibit this theorized failure precisely, we RL-finetune a language model and find evidence that our formal results are plausibly relevant in practice. We also propose a theoretical alternative that avoids this problem by replacing the "Don't do anything I wouldn't do" principle with "Don't do anything I mightn't do".

Large language models (LLMs) have demonstrated remarkable capabilities in generating high-quality texts across diverse domains. However, the potential misuse of LLMs has raised significant concerns, underscoring the urgent need for reliable detection of LLM-generated texts. Conventional training-based detectors often struggle with generalization, particularly in cross-domain and cross-model scenarios. In contrast, training-free methods, which focus on inherent discrepancies through carefully designed statistical features, offer improved generalization and interpretability. Despite this, existing training-free detection methods typically rely on global text sequence statistics, neglecting the modeling of local discriminative features, thereby limiting their detection efficacy. In this work, we introduce a novel training-free detector, termed \textbf{Lastde} that synergizes local and global statistics for enhanced detection. For the first time, we introduce time series analysis to LLM-generated text detection, capturing the temporal dynamics of token probability sequences. By integrating these local statistics with global ones, our detector reveals significant disparities between human and LLM-generated texts. We also propose an efficient alternative, \textbf{Lastde++} to enable real-time detection. Extensive experiments on six datasets involving cross-domain, cross-model, and cross-lingual detection scenarios, under both white-box and black-box settings, demonstrated that our method consistently achieves state-of-the-art performance. Furthermore, our approach exhibits greater robustness against paraphrasing attacks compared to existing baseline methods.

Fueled by the popularity of the transformer architecture in deep learning, several works have investigated what formal languages a transformer can learn. Nonetheless, existing results remain hard to compare and a fine-grained understanding of the trainability of transformers on regular languages is still lacking. We investigate transformers trained on regular languages from a mechanistic interpretability perspective. Using an extension of the $L^*$ algorithm, we extract Moore machines from transformers. We empirically find tighter lower bounds on the trainability of transformers, when a finite number of symbols determine the state. Additionally, our mechanistic insight allows us to characterise the regular languages a one-layer transformer can learn with good length generalisation. However, we also identify failure cases where the determining symbols get misrecognised due to saturation of the attention mechanism.