We separate LLM biases into deep biases that survive prompt reframing and shallow ones that do not. Across 4,442 opinion prompts and four models, only about a quarter of concentrated preferences persist under reframing, and these deep biases resist both fine-tuning and prompt-based debiasing.

We introduce the first benchmark measuring pattern-completion bias in screenshot-to-code generation, showing that frontier MLLMs complete familiar UI patterns instead of reading what is actually on screen.

We present FeatureSHAP, an interpretability framework for software engineering tasks that attributes Shapley scores to input features based on their contributions to model output.

We show that training LLMs on chain-of-thought reasoning improves their performance in sentiment analysis while enabling human-like explanations.

We show that an even split in the annotation budget between synthetic and human-curated data collection yields the best dataset for medical speech summarization LLMs.

We propose a mathematical theorem on the upper bound of the expressiveness of a neural network based on their geometric properties and apply it to model pruning.

We show that VLMs interpret UML diagrams from pretrained priors rather than the diagram itself: simply reversing relation arrows drops open-source model accuracy by 33.48% on average.

We present S-Chain, a medical dataset for strutured visual reasoning. Training medical models on S-Chain improves their accuracy and explainability.