We investigate how Large Language Models (LLMs) behave when simulating political discourse on social media. Leveraging 21 million interactions on X during the 2024 U.S. presidential election, we construct LLM agents based on 1,186 real users, prompting them to reply to politically salient tweets under controlled conditions. Agents are initialized either with minimal ideological cues (Zero Shot) or recent tweet history (Few Shot), allowing one-to-one comparisons with human replies. We evaluate three model families (Gemini, Mistral, and DeepSeek) across linguistic style, ideological consistency, and toxicity. We find that richer contextualization improves internal consistency but also amplifies polarization, stylized signals, and harmful language. We observe an emergent distortion that we call "generation exaggeration": a systematic amplification of salient traits beyond empirical baselines. Our analysis shows that LLMs do not emulate users, they reconstruct them. Their outputs, indeed, reflect internal optimization dynamics more than observed behavior, introducing structural biases that compromise their reliability as social proxies. This challenges their use in content moderation, deliberative simulations, and policy modeling.

Topic model and document-clustering evaluations either use automated metrics that align poorly with human preferences or require expert labels that are intractable to scale. We design a scalable human evaluation protocol and a corresponding automated approximation that reflect practitioners' real-world usage of models. Annotators -- or an LLM-based proxy -- review text items assigned to a topic or cluster, infer a category for the group, then apply that category to other documents. Using this protocol, we collect extensive crowdworker annotations of outputs from a diverse set of topic models on two datasets. We then use these annotations to validate automated proxies, finding that the best LLM proxies are statistically indistinguishable from a human annotator and can therefore serve as a reasonable substitute in automated evaluations. Package, web interface, and data are at https://github.com/ahoho/proxann

Semantic data representations are critical in artificial intelligence, significantly enhancing model performance in tasks like transfer and zero-shot learning (Lake et al., 2017). Central to this effort is to disentangle latent representations in generative models--representations where each latent dimension corresponds to an independent underlying factor of variation in the data. Disentanglement is achieved by leveraging statistical properties of the latent space and the dataset, enabling models where changes in one latent dimension affect only its corresponding factor without impacting others. This not only improves model interpretability but also enhances robustness against adversarial attacks (Yang et al., 2021). For a comprehensive review of disentanglement and its statistical underpinnings, see Wang et al. (2023). Datasets encountered in scientific research are often heterogeneous in modalities, fidelities, and accuracy where a particular entity or a state may be simultaneously associated with multiple images, graphs, vectors, scalar parameters, or labels with various associated measurement uncertainties.

This thesis contributes to the theoretical understanding of local update algorithms, especially Local SGD, in distributed and federated optimization under realistic models of data heterogeneity. A central focus is on the bounded second-order heterogeneity assumption, which is shown to be both necessary and sufficient for local updates to outperform centralized or mini-batch methods in convex and non-convex settings. The thesis establishes tight upper and lower bounds in several regimes for various local update algorithms and characterizes the min-max complexity of multiple problem classes. At its core is a fine-grained consensus-error-based analysis framework that yields sharper finite-time convergence bounds under third-order smoothness and relaxed heterogeneity assumptions. The thesis also extends to online federated learning, providing fundamental regret bounds under both first-order and bandit feedback. Together, these results clarify when and why local updates offer provable advantages, and the thesis serves as a self-contained guide for analyzing Local SGD in heterogeneous environments.

Water consumption is an increasingly critical dimension of computing sustainability, especially as AI workloads rapidly scale. However, current water impact assessment often overlooks where and when water stress is more severe. To fill in this gap, we present SCARF, the first general framework that evaluates water impact of computing by factoring in both spatial and temporal variations in water stress. SCARF calculates an Adjusted Water Impact (AWI) metric that considers both consumption volume and local water stress over time. Through three case studies on LLM serving, datacenters, and semiconductor fabrication plants, we show the hidden opportunities for reducing water impact by optimizing location and time choices, paving the way for water-sustainable computing. The code is available at https://github.com/jojacola/SCARF.

The visual world around us constantly evolves, from real-time news and social media trends to global infrastructure changes visible through satellite imagery and augmented reality enhancements. However, Multimodal Large Language Models (MLLMs), which automate many tasks, struggle to stay current, limited by the cutoff dates in their fixed training datasets. To quantify this stagnation, we introduce LiveVQA, the first-of-its-kind dataset featuring 107,143 samples and 12 categories data specifically designed to support research in both seeking and updating with live visual knowledge. Drawing from recent news articles, video platforms, and academic publications in April 2024-May 2025, LiveVQA enables evaluation of how models handle latest visual information beyond their knowledge boundaries and how current methods help to update them. Our comprehensive benchmarking of 17 state-of-the-art MLLMs reveals significant performance gaps on content beyond knowledge cutoff, and tool-use or agentic visual seeking framework drastically gain an average of 327% improvement. Furthermore, we explore parameter-efficient fine-tuning (PEFT) methods to update MLLMs with new visual knowledge. We dive deeply to the critical balance between adapter capacity and model capability when updating MLLMs with new visual knowledge. All the experimental dataset and source code are publicly available at: https://livevqa.github.io.

Leaderboards showcase the current capabilities and limitations of Large Language Models (LLMs). To motivate the development of LLMs that represent the linguistic and cultural diversity of the Spanish-speaking community, we present La Leaderboard, the first open-source leaderboard to evaluate generative LLMs in languages and language varieties of Spain and Latin America. La Leaderboard is a community-driven project that aims to establish an evaluation standard for everyone interested in developing LLMs for the Spanish-speaking community. This initial version combines 66 datasets in Basque, Catalan, Galician, and different Spanish varieties, showcasing the evaluation results of 50 models. To encourage community-driven development of leaderboards in other languages, we explain our methodology, including guidance on selecting the most suitable evaluation setup for each downstream task. In particular, we provide a rationale for using fewer few-shot examples than typically found in the literature, aiming to reduce environmental impact and facilitate access to reproducible results for a broader research community.

Accurate molecular property predictions require 3D geometries, which are typically obtained using expensive methods such as density functional theory (DFT). Here, we attempt to obtain molecular geometries by relying solely on machine learning interatomic potential (MLIP) models. To this end, we first curate a large-scale molecular relaxation dataset comprising 3.5 million molecules and 300 million snapshots. Then MLIP foundation models are trained with supervised learning to predict energy and forces given 3D molecular structures. Once trained, we show that the foundation models can be used in different ways to obtain geometries either explicitly or implicitly. First, it can be used to obtain low-energy 3D geometries via geometry optimization, providing relaxed 3D geometries for downstream molecular property predictions. To mitigate potential biases and enhance downstream predictions, we introduce geometry fine-tuning based on the relaxed 3D geometries. Second, the foundation models can be directly fine-tuned for property prediction when ground truth 3D geometries are available. Our results demonstrate that MLIP foundation models trained on relaxation data can provide valuable molecular geometries that benefit property predictions.

In the science of emotion, it is widely assumed that folk emotion categories form a biological and psychological typology, and studies are routinely designed and analyzed to identify emotion-specific patterns. This approach shapes the observations that studies report, ultimately reinforcing the assumption that guided the investigation. Here, we reanalyzed data from one such typologically-guided study that reported mappings between individual brain patterns and group-averaged ratings of 34 emotion categories. Our reanalysis was guided by an alternative view of emotion categories as populations of variable, situated instances, and which predicts a priori that there will be significant variation in brain patterns within a category across instances. Correspondingly, our analysis made minimal assumptions about the structure of the variance present in the data. As predicted, we did not observe the original mappings and instead observed significant variation across individuals. These findings demonstrate how starting assumptions can ultimately impact scientific conclusions and suggest that a hypothesis must be supported using multiple analytic methods before it is taken seriously.

Recent advances in word embeddings and language models use large-scale, unlabelled data and self-supervised learning to boost NLP performance. Multilingual models, often trained on web-sourced data like Wikipedia, face challenges: few low-resource languages are included, their data is often noisy, and lack of labeled datasets makes it hard to evaluate performance outside high-resource languages like English. In this dissertation, we focus on languages spoken in Sub-Saharan Africa where all the indigenous languages in this region can be regarded as low-resourced in terms of the availability of labelled data for NLP tasks and unlabelled data found on the web. We analyse the noise in the publicly available corpora, and curate a high-quality corpus, demonstrating that the quality of semantic representations learned in word embeddings does not only depend on the amount of data but on the quality of pre-training data. We demonstrate empirically the limitations of word embeddings, and the opportunities the multilingual pre-trained language model (PLM) offers especially for languages unseen during pre-training and low-resource scenarios. We further study how to adapt and specialize multilingual PLMs to unseen African languages using a small amount of monolingual texts. To address the under-representation of the African languages in NLP research, we developed large scale human-annotated labelled datasets for 21 African languages in two impactful NLP tasks: named entity recognition and machine translation. We conduct an extensive empirical evaluation using state-of-the-art methods across supervised, weakly-supervised, and transfer learning settings.