Large language models (LLMs) frequently generate confident yet inaccurate responses, introducing significant risks for deployment in safety-critical domains. We present a novel approach to detecting model hallucination through systematic analysis of information flow across model layers when processing inputs with insufficient or ambiguous context. Our investigation reveals that hallucination manifests as usable information deficiencies in inter-layer transmissions. While existing approaches primarily focus on final-layer output analysis, we demonstrate that tracking cross-layer information dynamics ($\mathcal{L}$I) provides robust indicators of model reliability, accounting for both information gain and loss during computation. $\mathcal{L}$I improves model reliability by immediately integrating with universal LLMs without additional training or architectural modifications.

Internet connectivity in schools is critical to provide students with the digital literary skills necessary to compete in modern economies. In order for governments to effectively implement digital infrastructure development in schools, accurate internet connectivity information is required. However, traditional survey-based methods can exceed the financial and capacity limits of governments. Open-source Earth Observation (EO) datasets have unlocked our ability to observe and understand socio-economic conditions on Earth from space, and in combination with Machine Learning (ML), can provide the tools to circumvent costly ground-based survey methods to support infrastructure development. In this paper, we present our work on school internet connectivity prediction using EO and ML. We detail the creation of our multi-modal, freely-available satellite imagery and survey information dataset, leverage the latest geographically-aware location encoders, and introduce the first results of using the new European Space Agency phi-lab geographically-aware foundational model to predict internet connectivity in Botswana and Rwanda. We find that ML with EO and ground-based auxiliary data yields the best performance in both countries, for accuracy, F1 score, and False Positive rates, and highlight the challenges of internet connectivity prediction from space with a case study in Kigali, Rwanda. Our work showcases a practical approach to support data-driven digital infrastructure development in low-resource settings, leveraging freely available information, and provide cleaned and labelled datasets for future studies to the community through a unique collaboration between UNICEF and the European Space Agency phi-lab.

The proliferation of disinformation presents a growing threat to societal trust and democracy, necessitating robust and scalable Fact-Checking systems. In this work, we present Dynamic Evidence-based FAct-checking with Multimodal Experts (DEFAME), a modular, zero-shot MLLM pipeline for open-domain, text-image claim verification. DEFAME frames the problem of fact-checking as a six-stage process, dynamically deciding about the usage of external tools for the retrieval of textual and visual evidence. In addition to the claim's veracity, DEFAME returns a justification accompanied by a comprehensive, multimodal fact-checking report. While most alternatives either focus on sub-tasks of fact-checking, lack explainability or are limited to text-only inputs, DEFAME solves the problem of fact-checking end-to-end, including claims with images or those that require visual evidence. Evaluation on the popular benchmarks VERITE, AVeriTeC, and MOCHEG shows that DEFAME surpasses all previous methods, establishing it as the new state-of-the-art fact-checking system.

Many algorithms have been developed to estimate probability distributions subject to differential privacy (DP): such an algorithm takes as input independent samples from a distribution and estimates the density function in a way that is insensitive to any one sample. A recent line of work, initiated by Raskhodnikova et al. (Neurips '21), explores a weaker objective: a differentially private algorithm that approximates a single sample from the distribution. Raskhodnikova et al. studied the sample complexity of DP \emph{single-sampling} i.e., the minimum number of samples needed to perform this task. They showed that the sample complexity of DP single-sampling is less than the sample complexity of DP learning for certain distribution classes. We define two variants of \emph{multi-sampling}, where the goal is to privately approximate $m>1$ samples. This better models the realistic scenario where synthetic data is needed for exploratory data analysis. A baseline solution to \emph{multi-sampling} is to invoke a single-sampling algorithm $m$ times on independently drawn datasets of samples. When the data comes from a finite domain, we improve over the baseline by a factor of $m$ in the sample complexity. When the data comes from a Gaussian, Ghazi et al. (Neurips '23) show that \emph{single-sampling} can be performed under approximate differential privacy; we show it is possible to \emph{single- and multi-sample Gaussians with known covariance subject to pure DP}. Our solution uses a variant of the Laplace mechanism that is of independent interest. We also give sample complexity lower bounds, one for strong multi-sampling of finite distributions and another for weak multi-sampling of bounded-covariance Gaussians.

Language models (LMs) have excelled in various broad domains. However, to ensure their safe and effective integration into real-world educational settings, they must demonstrate proficiency in specific, granular areas of knowledge. Existing cloze-style benchmarks, commonly used to evaluate LMs' knowledge, have three major limitations. They: 1) do not cover the educational domain; 2) typically focus on low-complexity, generic knowledge or broad domains, which do not adequately assess the models' knowledge in specific subjects; and 3) often rely on templates that can bias model predictions. Here, we introduce MALAMUTE, a multilingual, template-free, and highly granular probing dataset comprising expert-written, peer-reviewed probes from 71 university-level textbooks across three languages (English, Spanish, and Polish). MALAMUTE is the first education-based cloze-style dataset. It covers eight domains, each with up to 14 subdomains, further broken down into concepts and concept-based prompts, totaling 33,361 university curriculum concepts and 116,887 prompts. MALAMUTE's fine granularity, educational focus, and inclusion of both sentence-level and paragraph-level prompts make it an ideal tool for evaluating LMs' course-related knowledge. Our evaluation of masked and causal LMs on MALAMUTE shows that despite overall proficiency, they have significant gaps in knowledge when examined closely on specific subjects, hindering their safe use in classrooms and underscoring the need for further development.

A myriad of measures to illustrate performance of predictive artificial intelligence (AI) models have been proposed in the literature. Selecting appropriate performance measures is essential for predictive AI models that are developed to be used in medical practice, because poorly performing models may harm patients and lead to increased costs. We aim to assess the merits of classic and contemporary performance measures when validating predictive AI models for use in medical practice. We focus on models with a binary outcome. We discuss 32 performance measures covering five performance domains (discrimination, calibration, overall, classification, and clinical utility) along with accompanying graphical assessments. The first four domains cover statistical performance, the fifth domain covers decision-analytic performance. We explain why two key characteristics are important when selecting which performance measures to assess: (1) whether the measure's expected value is optimized when it is calculated using the correct probabilities (i.e., a "proper" measure), and (2) whether they reflect either purely statistical performance or decision-analytic performance by properly considering misclassification costs. Seventeen measures exhibit both characteristics, fourteen measures exhibited one characteristic, and one measure possessed neither characteristic (the F1 measure). All classification measures (such as classification accuracy and F1) are improper for clinically relevant decision thresholds other than 0.5 or the prevalence. We recommend the following measures and plots as essential to report: AUROC, calibration plot, a clinical utility measure such as net benefit with decision curve analysis, and a plot with probability distributions per outcome category.

We consider the problem of releasing a sparse histogram under $(\varepsilon, \delta)$-differential privacy. The stability histogram independently adds noise from a Laplace or Gaussian distribution to the non-zero entries and removes those noisy counts below a threshold. Thereby, the introduction of new non-zero values between neighboring histograms is only revealed with probability at most $\delta$, and typically, the value of the threshold dominates the error of the mechanism. We consider the variant of the stability histogram with Gaussian noise. Recent works ([Joseph and Yu, COLT '24] and [Lebeda, SOSA '25]) reduced the error for private histograms using correlated Gaussian noise. However, these techniques can not be directly applied in the very sparse setting. Instead, we adopt Lebeda's technique and show that adding correlated noise to the non-zero counts only allows us to reduce the magnitude of noise when we have a sparsity bound. This, in turn, allows us to use a lower threshold by up to a factor of $1/2$ compared to the non-correlated noise mechanism. We then extend our mechanism to a setting without a known bound on sparsity. Additionally, we show that correlated noise can give a similar improvement for the more practical discrete Gaussian mechanism.

There is a growing consensus that information is the "lifeblood of good governance" (Kolt et al., 2024) and that information sharing should be one of the "natural initial target[s]" of AI governance (Bommasani et al., 2024). Up-to-date and reliable information about AI systems' capabilities and how capabilities will develop in the future can help developers, governments, and researchers advance safety evaluations (Frontier Model Forum, 2024), develop best practices (UK DSIT, 2023), and respond effectively to the new risks posed by frontier AI (Kolt et al., 2024). Information sharing also supports regulatory visibility (Anderljung et al., 2023) and can thus enable better-informed AI governance (O'Brien et al., 2024). Further, access to knowledge about AI systems' potential risks allows AI systems claims to be scrutinized more effectively (Brundage et al., 2020). By contrast, information asymmetries could lead regulators to miscalibrated over-regulation--or under-regulation--of AI (Ball & Kokotajlo, 2024) and could contribute to the "pacing problem," a situation in which government oversight consistently lags behind technology development (Marchant et al., 2011). In short, there is a strong case for information sharing being one "key to making AI go well" (Ball & Kokotajlo, 2024). The Frontier AI Safety Commitments ("FAISC") are an important step towards more comprehensive information sharing by AI developers.

Estimating the expected output quality of generation systems is central to NLG. This paper qualifies the notion that automatic metrics are not as good as humans in estimating system-level quality. Statistically, humans are unbiased, high variance estimators, while metrics are biased, low variance estimators. We compare these estimators by their error in pairwise prediction (which generation system is better?) using the bootstrap. Measuring this error is complicated: predictions are evaluated against noisy, human predicted labels instead of the ground truth, and metric predictions fluctuate based on the test sets they were calculated on. By applying a bias-variance-noise decomposition, we adjust this error to a noise-free, infinite test set setting. Our analysis compares the adjusted error of metrics to humans and a derived, perfect segment-level annotator, both of which are unbiased estimators dependent on the number of judgments collected. In MT, we identify two settings where metrics outperform humans due to a statistical advantage in variance: when the number of human judgments used is small, and when the quality difference between compared systems is small. The data and code to reproduce our analyses are available at https://github.com/johntzwei/metric-statistical-advantage .

We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the large-scale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only 0.6B parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality.