Global Station Weather Forecasting (GSWF) is crucial for various sectors, including aviation, agriculture, energy, and disaster preparedness. Recent advancements in deep learning have significantly improved the accuracy of weather predictions by optimizing models based on public meteorological data. However, existing public datasets for GSWF optimization and benchmarking still suffer from significant limitations, such as small sizes, limited temporal coverage, and a lack of comprehensive variables. These shortcomings prevent them from effectively reflecting the benchmarks of current forecasting methods and fail to support the real needs of operational weather forecasting. To address these challenges, we present the WEATHER-5K dataset. This dataset comprises a comprehensive collection of data from 5,672 weather stations worldwide, spanning a 10-year period with one-hour intervals. It includes multiple crucial weather elements, providing a more reliable and interpretable resource for forecasting. Furthermore, our WEATHER-5K dataset can serve as a benchmark for comprehensively evaluating existing well-known forecasting models, extending beyond GSWF methods to support future time-series research challenges and opportunities.

Frontier AI systems are making transformative impacts across society, but such benefits are not without costs: models trained on web-scale datasets containing personal and private data raise profound concerns about data privacy and security. Language models are trained on extensive corpora including potentially sensitive or proprietary information, and the risk of data leakage -- where the model response reveals pieces of such information -- remains inadequately understood. Prior work has investigated what factors drive memorization and have identified that sequence complexity and the number of repetitions drive memorization. Here, we focus on the evolution of memorization over training. We begin by reproducing findings that the probability of memorizing a sequence scales logarithmically with the number of times it is present in the data. We next show that sequences which are apparently not memorized after the first encounter can be "uncovered" throughout the course of training even without subsequent encounters, a phenomenon we term "latent memorization". The presence of latent memorization presents a challenge for data privacy as memorized sequences may be hidden at the final checkpoint of the model but remain easily recoverable. To this end, we develop a diagnostic test relying on the cross entropy loss to uncover latent memorized sequences with high accuracy.

Socio-demographic prompting is a commonly employed approach to study cultural biases in LLMs as well as for aligning models to certain cultures. In this paper, we systematically probe four LLMs (Llama 3, Mistral v0.2, GPT-3.5 Turbo and GPT-4) with prompts that are conditioned on culturally sensitive and non-sensitive cues, on datasets that are supposed to be culturally sensitive (EtiCor and CALI) or neutral (MMLU and ETHICS). We observe that all models except GPT-4 show significant variations in their responses on both kinds of datasets for both kinds of prompts, casting doubt on the robustness of the culturally-conditioned prompting as a method for eliciting cultural bias in models or as an alignment strategy. The work also calls rethinking the control experiment design to tease apart the cultural conditioning of responses from "placebo effect", i.e., random perturbations of model responses due to arbitrary tokens in the prompt.

Large vision-language models (LVLMs) have recently dramatically pushed the state of the art in image captioning and many image understanding tasks (e.g., visual question answering). LVLMs, however, often \textit{hallucinate} and produce captions that mention concepts that cannot be found in the image. These hallucinations erode the trustworthiness of LVLMs and are arguably among the main obstacles to their ubiquitous adoption. Recent work suggests that addition of grounding objectives -- those that explicitly align image regions or objects to text spans -- reduces the amount of LVLM hallucination. Although intuitive, this claim is not empirically justified as the reduction effects have been established, we argue, with flawed evaluation protocols that (i) rely on data (i.e., MSCOCO) that has been extensively used in LVLM training and (ii) measure hallucination via question answering rather than open-ended caption generation. In this work, in contrast, we offer the first systematic analysis of the effect of fine-grained object grounding on LVLM hallucination under an evaluation protocol that more realistically captures LVLM hallucination in open generation. Our extensive experiments over three backbone LLMs reveal that grounding objectives have little to no effect on object hallucination in open caption generation.

Large Language Models (LLMs) are routinely used in retrieval-augmented applications to orchestrate tasks and process inputs from users and other sources. These inputs, even in a single LLM interaction, can come from a variety of sources, of varying trustworthiness and provenance. This opens the door to prompt injection attacks, where the LLM receives and acts upon instructions from supposedly data-only sources, thus deviating from the user's original instructions. We define this as task drift, and we propose to catch it by scanning and analyzing the LLM's activations. We compare the LLM's activations before and after processing the external input in order to detect whether this input caused instruction drift. We develop two probing methods and find that simply using a linear classifier can detect drift with near perfect ROC AUC on an out-of-distribution test set. We show that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Our setup does not require any modification of the LLM (e.g., fine-tuning) or any text generation, thus maximizing deployability and cost efficiency and avoiding reliance on unreliable model output. To foster future research on activation-based task inspection, decoding, and interpretability, we will release our large-scale TaskTracker toolkit, comprising a dataset of over 500K instances, representations from 4 SoTA language models, and inspection tools.

Label scarcity problem is the main challenge that hinders the wide application of deep learning systems in automatic cardiovascular diseases (CVDs) detection using electrocardiography (ECG). Tuning pre-trained models alleviates this problem by transferring knowledge learned from large datasets to downstream small datasets. However, bottlenecks in computational efficiency and CVDs detection performance limit its clinical applications. It is difficult to improve the detection performance without significantly sacrificing model computational efficiency. Here, we propose a computation-efficient semi-supervised learning paradigm (FastECG) for robust and computation-efficient CVDs detection using ECG. It enables a robust adaptation of pre-trained models on downstream datasets with limited supervision and high computational efficiency. First, a random-deactivation technique is developed to achieve robust and fast low-rank adaptation of pre-trained weights. Subsequently, we propose a one-shot rank allocation module to determine the optimal ranks for the update matrices of the pre-trained weights. Finally, a lightweight semi-supervised learning pipeline is introduced to enhance model performance by leveraging labeled and unlabeled data with high computational efficiency. Extensive experiments on four downstream ECG datasets demonstrate that FastECG not only outperforms the state-of-the-art methods in multi-label CVDs detection but also consumes fewer GPU footprints, training time, and parameter storage space. As such, this paradigm provides an effective solution for achieving high computational efficiency and robust detection performance in the clinical applications of pre-trained models under limited supervision.

We introduce eigenpruning, a method that removes singular values from weight matrices in an LLM to improve its performance in a particular task. This method is inspired by interpretability methods designed to automatically find subnetworks of a model which solve a specific task. In our tests, the pruned model outperforms the original model by a large margin, while only requiring minimal computation to prune the weight matrices. In the case of a small synthetic task in integer multiplication, the Phi-2 model can improve its accuracy in the test set from 13.75% to 97.50%. Interestingly, these results seem to indicate the existence of a computation path that can solve the task very effectively, but it was not being used by the original model. Finally, we publicly release our implementation.

Retrieval Augmented Generation (RAG) is widely employed to ground responses to queries on domain-specific documents. But do RAG implementations leave out important information or excessively include irrelevant information? To allay these concerns, it is necessary to annotate domain-specific benchmarks to evaluate information retrieval (IR) performance, as relevance definitions vary across queries and domains. Furthermore, such benchmarks should be cost-efficiently annotated to avoid annotation selection bias. In this paper, we propose DIRAS (Domain-specific Information Retrieval Annotation with Scalability), a manual-annotation-free schema that fine-tunes open-sourced LLMs to annotate relevance labels with calibrated relevance probabilities. Extensive evaluation shows that DIRAS fine-tuned models achieve GPT-4-level performance on annotating and ranking unseen (query, document) pairs, and is helpful for real-world RAG development.

One way to address safety risks from large language models (LLMs) is to censor dangerous knowledge from their training data. While this removes the explicit information, implicit information can remain scattered across various training documents. Could an LLM infer the censored knowledge by piecing together these implicit hints? As a step towards answering this question, we study inductive out-of-context reasoning (OOCR), a type of generalization in which LLMs infer latent information from evidence distributed across training documents and apply it to downstream tasks without in-context learning. Using a suite of five tasks, we demonstrate that frontier LLMs can perform inductive OOCR. In one experiment we finetune an LLM on a corpus consisting only of distances between an unknown city and other known cities. Remarkably, without in-context examples or Chain of Thought, the LLM can verbalize that the unknown city is Paris and use this fact to answer downstream questions. Further experiments show that LLMs trained only on individual coin flip outcomes can verbalize whether the coin is biased, and those trained only on pairs $(x,f(x))$ can articulate a definition of $f$ and compute inverses. While OOCR succeeds in a range of cases, we also show that it is unreliable, particularly for smaller LLMs learning complex structures. Overall, the ability of LLMs to "connect the dots" without explicit in-context learning poses a potential obstacle to monitoring and controlling the knowledge acquired by LLMs.

Lexical ambiguity -- where a single wordform takes on distinct, context-dependent meanings -- serves as a useful tool to compare across different large language models' (LLMs') ability to form distinct, contextualized representations of the same stimulus. Few studies have systematically compared LLMs' contextualized word embeddings for languages beyond English. Here, we evaluate multiple bidirectional transformers' (BERTs') semantic representations of Spanish ambiguous nouns in context. We develop a novel dataset of minimal-pair sentences evoking the same or different sense for a target ambiguous noun. In a pre-registered study, we collect contextualized human relatedness judgments for each sentence pair. We find that various BERT-based LLMs' contextualized semantic representations capture some variance in human judgments but fall short of the human benchmark, and for Spanish -- unlike English -- model scale is uncorrelated with performance. We also identify stereotyped trajectories of target noun disambiguation as a proportion of traversal through a given LLM family's architecture, which we partially replicate in English. We contribute (1) a dataset of controlled, Spanish sentence stimuli with human relatedness norms, and (2) to our evolving understanding of the impact that LLM specification (architectures, training protocols) exerts on contextualized embeddings.