Recent progress in self-supervision has shown that pre-training large neural networks on vast amounts of unsupervised data can lead to substantial increases in generalization to downstream tasks. Such models, recently coined foundation models, have been transformational to the field of natural language processing. Variants have also been proposed for image data, but their applicability to remote sensing tasks is limited. To stimulate the development of foundation models for Earth monitoring, we propose a benchmark comprised of six classification and six segmentation tasks, which were carefully curated and adapted to be both relevant to the field and well-suited for model evaluation. We accompany this benchmark with a robust methodology for evaluating models and reporting aggregated results to enable a reliable assessment of progress. Finally, we report results for 20 baselines to gain information about the performance of existing models. We believe that this benchmark will be a driver of progress across a variety of Earth monitoring tasks.

We propose a physics-informed machine learning method for uncertainty quantification in high-dimensional inverse problems. In this method, the states and parameters of partial differential equations (PDEs) are approximated with truncated conditional Karhunen-Lo\`eve expansions (CKLEs), which, by construction, match the measurements of the respective variables. The maximum a posteriori (MAP) solution of the inverse problem is formulated as a minimization problem over CKLE coefficients where the loss function is the sum of the norm of PDE residuals and the $\ell_2$ regularization term. This MAP formulation is known as the physics-informed CKLE (PICKLE) method. Uncertainty in the inverse solution is quantified in terms of the posterior distribution of CKLE coefficients, and we sample the posterior by solving a randomized PICKLE minimization problem, formulated by adding zero-mean Gaussian perturbations in the PICKLE loss function. We call the proposed approach the randomized PICKLE (rPICKLE) method. For linear and low-dimensional nonlinear problems (15 CKLE parameters), we show analytically and through comparison with Hamiltonian Monte Carlo (HMC) that the rPICKLE posterior converges to the true posterior given by the Bayes rule. For high-dimensional non-linear problems with 2000 CKLE parameters, we numerically demonstrate that rPICKLE posteriors are highly informative--they provide mean estimates with an accuracy comparable to the estimates given by the MAP solution and the confidence interval that mostly covers the reference solution. We are not able to obtain the HMC posterior to validate rPICKLE's convergence to the true posterior due to the HMC's prohibitive computational cost for the considered high-dimensional problems. Our results demonstrate the advantages of rPICKLE over HMC for approximately sampling high-dimensional posterior distributions subject to physics constraints.

Gif-sur-Yvette 91191 Abstract We suggest the implementation of the Dual Use Research of Concern (DURC) framework, originally designed for life sciences, to the domain of generative AI, with a specific focus on Large Language Models (LLMs). With its demonstrated advantages and drawbacks in biological research, we believe the DURC criteria can be effectively redefined for LLMs, potentially contributing to improved AI governance. Acknowledging the balance that must be struck when employing the DURC framework, we highlight its crucial political role in enhancing societal awareness of the impact of generative AI. As a final point, we offer a series of specific recommendations for applying the DURC approach to LLM research. Keywords: Dual Use Research of Concern (DURC), Generative AI, Large Language Models (LLMs), AI Ethics Conflict of interest No conflict of interest to report. Funding This research was supported through projects TechEthos (grant number 101006249) and MultiRATE (grant number 101073929) funded by the European Commission Horizon program. Ethics approval No human subjects were involved in the study. Consent No data needing consent has been used. Data availability statement In this article, we do not analyze or generate any datasets. Author Contribution All authors contributed to the study conception and design. Sections 1 and 4 were written with equal contribution. Sections 2 and 3 were conceived by Adomaitis and later edited by Grinbaum.

The advent of large language models (LLMs) has heightened interest in their potential for multimodal applications that integrate language and vision. This paper explores the capabilities of GPT-4V in the realms of geography, environmental science, agriculture, and urban planning by evaluating its performance across a variety of tasks. Data sources comprise satellite imagery, aerial photos, ground-level images, field images, and public datasets. The model is evaluated on a series of tasks including geo-localization, textual data extraction from maps, remote sensing image classification, visual question answering, crop type identification, disease/pest/weed recognition, chicken behavior analysis, agricultural object counting, urban planning knowledge question answering, and plan generation. The results indicate the potential of GPT-4V in geo-localization, land cover classification, visual question answering, and basic image understanding. However, there are limitations in several tasks requiring fine-grained recognition and precise counting. While zero-shot learning shows promise, performance varies across problem domains and image complexities. The work provides novel insights into GPT-4V's capabilities and limitations for real-world geospatial, environmental, agricultural, and urban planning challenges. Further research should focus on augmenting the model's knowledge and reasoning for specialized domains through expanded training. Overall, the analysis demonstrates foundational multimodal intelligence, highlighting the potential of multimodal foundation models (FMs) to advance interdisciplinary applications at the nexus of computer vision and language.

The advances of Large Language Models (LLMs) are expanding their utility in both academic research and practical applications. Recent social science research has explored the use of these "black-box" LLM agents for simulating complex social systems and potentially substituting human subjects in experiments. Our study delves into this emerging domain, investigating the extent to which LLMs exhibit key social interaction principles, such as social learning, social preference, and cooperative behavior, in their interactions with humans and other agents. We develop a novel framework for our study, wherein classical laboratory experiments involving human subjects are adapted to use LLM agents. This approach involves step-by-step reasoning that mirrors human cognitive processes and zero-shot learning to assess the innate preferences of LLMs. Our analysis of LLM agents' behavior includes both the primary effects and an in-depth examination of the underlying mechanisms. Focusing on GPT-4, the state-of-the-art LLM, our analyses suggest that LLM agents appear to exhibit a range of human-like social behaviors such as distributional and reciprocity preferences, responsiveness to group identity cues, engagement in indirect reciprocity, and social learning capabilities. However, our analysis also reveals notable differences: LLMs demonstrate a pronounced fairness preference, weaker positive reciprocity, and a more calculating approach in social learning compared to humans. These insights indicate that while LLMs hold great promise for applications in social science research, such as in laboratory experiments and agent-based modeling, the subtle behavioral differences between LLM agents and humans warrant further investigation. Careful examination and development of protocols in evaluating the social behaviors of LLMs are necessary before directly applying these models to emulate human behavior.

Multi-hop question answering (MQA) is one of the challenging tasks to evaluate machine's comprehension and reasoning abilities, where large language models (LLMs) have widely achieved the human-comparable performance. Due to the dynamics of knowledge facts in real world, knowledge editing has been explored to update model with the up-to-date facts while avoiding expensive re-training or fine-tuning. Starting from the edited fact, the updated model needs to provide cascading changes in the chain of MQA. The previous art simply adopts a mix-up prompt to instruct LLMs conducting multiple reasoning tasks sequentially, including question decomposition, answer generation, and conflict checking via comparing with edited facts. However, the coupling of these functionally-diverse reasoning tasks inhibits LLMs' advantages in comprehending and answering questions while disturbing them with the unskilled task of conflict checking. We thus propose a framework, Programmable knowledge editing for Multi-hop Question Answering (PokeMQA), to decouple the jobs. Specifically, we prompt LLMs to decompose knowledge-augmented multi-hop question, while interacting with a detached trainable scope detector to modulate LLMs behavior depending on external conflict signal. The experiments on three LLM backbones and two benchmark datasets validate our superiority in knowledge editing of MQA, outperforming all competitors by a large margin in almost all settings and consistently producing reliable reasoning process.

The capacity to generalize to future unseen data stands as one of the utmost crucial attributes of deep neural networks. Sharpness-Aware Minimization (SAM) aims to enhance the generalizability by minimizing worst-case loss using one-step gradient ascent as an approximation. However, as training progresses, the non-linearity of the loss landscape increases, rendering one-step gradient ascent less effective. On the other hand, multi-step gradient ascent will incur higher training cost. In this paper, we introduce a normalized Hessian trace to accurately measure the curvature of loss landscape on {\em both} training and test sets. In particular, to counter excessive non-linearity of loss landscape, we propose Curvature Regularized SAM (CR-SAM), integrating the normalized Hessian trace as a SAM regularizer. Additionally, we present an efficient way to compute the trace via finite differences with parallelism. Our theoretical analysis based on PAC-Bayes bounds establishes the regularizer's efficacy in reducing generalization error. Empirical evaluation on CIFAR and ImageNet datasets shows that CR-SAM consistently enhances classification performance for ResNet and Vision Transformer (ViT) models across various datasets. Our code is available at https://github.com/TrustAIoT/CR-SAM.

User modeling (UM) aims to discover patterns or learn representations from user data about the characteristics of a specific user, such as profile, preference, and personality. The user models enable personalization and suspiciousness detection in many online applications such as recommendation, education, and healthcare. Two common types of user data are text and graph, as the data usually contain a large amount of user-generated content (UGC) and online interactions. The research of text and graph mining is developing rapidly, contributing many notable solutions in the past two decades. Recently, large language models (LLMs) have shown superior performance on generating, understanding, and even reasoning over text data. The approaches of user modeling have been equipped with LLMs and soon become outstanding. This article summarizes existing research about how and why LLMs are great tools of modeling and understanding UGC. Then it reviews a few categories of large language models for user modeling (LLM-UM) approaches that integrate the LLMs with text and graph-based methods in different ways. Then it introduces specific LLM-UM techniques for a variety of UM applications. Finally, it presents remaining challenges and future directions in the LLM-UM research. We maintain the reading list at: https://github.com/TamSiuhin/LLM-UM-Reading

Adversarial transferability enables black-box attacks on unknown victim deep neural networks (DNNs), rendering attacks viable in real-world scenarios. Current transferable attacks create adversarial perturbation over the entire image, resulting in excessive noise that overfit the source model. Concentrating perturbation to dominant image regions that are model-agnostic is crucial to improving adversarial efficacy. However, limiting perturbation to local regions in the spatial domain proves inadequate in augmenting transferability. To this end, we propose a transferable adversarial attack with fine-grained perturbation optimization in the frequency domain, creating centralized perturbation. We devise a systematic pipeline to dynamically constrain perturbation optimization to dominant frequency coefficients. The constraint is optimized in parallel at each iteration, ensuring the directional alignment of perturbation optimization with model prediction. Our approach allows us to centralize perturbation towards sample-specific important frequency features, which are shared by DNNs, effectively mitigating source model overfitting. Experiments demonstrate that by dynamically centralizing perturbation on dominating frequency coefficients, crafted adversarial examples exhibit stronger transferability, and allowing them to bypass various defenses.

This paper conducts a robustness audit of the safety feedback of PaLM 2 through a novel toxicity rabbit hole framework introduced here. Starting with a stereotype, the framework instructs PaLM 2 to generate more toxic content than the stereotype. Every subsequent iteration it continues instructing PaLM 2 to generate more toxic content than the previous iteration until PaLM 2 safety guardrails throw a safety violation. Our experiments uncover highly disturbing antisemitic, Islamophobic, racist, homophobic, and misogynistic (to list a few) generated content that PaLM 2 safety guardrails do not evaluate as highly unsafe. We briefly discuss the generalizability of this framework across eight other large language models.