A study of activation steering across 50 behaviors, examining how effectiveness varies by behavior type. Stronger steering does not always improve control, and separation between steering vectors does not predict success.
A Ukrainian multimodal benchmark with over 4,300 questions across 12 academic disciplines, alongside evaluations of image captioning, visual question answering, and cultural knowledge.
An evaluation of gender bias in five language models, comparing their predictions about occupational roles with human perceptions, workforce statistics, and a gender-neutral baseline.
Combining hypothetical document embeddings with retrieval decisions and passage critiques, evaluated on SQuAD questions.
Evaluating generative augmentation for low-resource Ukrainian handwriting recognition, examining whether VAE- and GAN-generated samples improve downstream classification and where data scarcity limits their effectiveness.
A conversational workflow for clarifying ambiguous questions, retrieving official pension information, and routing requests between document-grounded responses and calculation tools.
Developing LSTM models of symptom sequences and XGBoost feature analysis within a symptom-tracking prototype, using synthetic data to explore the prediction pipeline.
Comparing centrality-based predictions with simulated traffic in Berlin and San Francisco to examine how road-network structure relates to emerging congestion.
Modeling fire spread under varying environmental conditions and comparing targeted, uniform, and random tree-removal strategies through repeated simulations.
Comparing Gaussian, Student-t, and outlier-mixture regression models for San Francisco temperature trends using posterior predictive checks and leave-one-out cross-validation.
Estimating missing GP visit counts across UK regions and age groups, comparing complete and partial pooling with zero-inflated Poisson models and predictive model evaluation.