Research

Our research develops methods for understanding, evaluating, and improving language technologies. Recent work spans large language models, data-centric machine learning, model interpretability, and language generation, often in collaboration with researchers in other fields.

1. Large language models: behavior, reasoning, and evaluation

We study how language models use examples, represent information, and make mistakes, with an eye toward evaluations that reveal more than a single score. Recent work investigates consistency in mathematical reasoning, syntactic causes of mathematical errors, and the influence of data in in-context learning. We also examine how model behavior affects practical settings such as automated grading.

• Consistency of LLMs to Comparative Statements in Mathematical Reasoning Tasks (2026)
• Syntactic Blind Spots: How Misalignment Leads to LLMs’ Mathematical Errors (2025)
• The Good, the Bad, and the Debatable: A Survey on the Impacts of Data for In-Context Learning (2025)
• In-Context Learning (and Unlearning) of Length Biases (2025)
• Evaluating the Effectiveness of Language Models for Discrete Mathematics Grading (2026)

2. Interpretability, fairness, and model behavior

We design methods to understand what models rely on and how their internal representations can be analyzed or controlled. This includes explanations of model predictions, sensitivity and bias audits, and techniques for examining or steering concepts in language models.

• CASE: Contrastive Activation for Class-Sensitive Explanations (2026)
• White-Box Sensitivity Auditing with Steering Vectors (2026)
• Unsupervised Concept Vector Extraction for Bias Control in LLMs (2025)
• Addressing Both Statistical and Causal Gender Fairness in NLP Models (2024)
• REV: Information-Theoretic Evaluation of Free-Text Rationales (2023)

3. Data-centric machine learning

Data quality, selection, and attribution are central to reliable model development. We study how to identify useful training examples, measure their contributions, and make data valuation practical, private, and robust—including for large language model training.

• PrAM: Offline Checkpoint Selection for Post-Hoc Training Data Attribution (2026)
• Selectformer: Private and Practical Data Selection for Transformers (2025)
• Data Selection for Fine-tuning Large Language Models Using Transferred Shapley Values (2023)
• Efficient NLP Model Finetuning via Data Filtering (2023)
• Secure and Effective Data Appraisal for Machine Learning (2023)

4. Language generation, summarization, and datasets

We build and evaluate systems and resources for generating and understanding language. Recent projects explore aspect-based summarization, knowledge-grounded dialogue, and datasets that support story understanding and structured information extraction.

• Improving Aspect-Based Summarization via Contrastive Learning with Anchored Negative Examples (2025)
• Synopses of Movie Narratives: a Video-Language Dataset for Story Understanding (2025)
• PLAtE: A Large-scale Dataset for List Page Web Extraction (2023)
• Blending Reward Functions via Few Expert Demonstrations for Faithful and Accurate Knowledge-Grounded Dialogue Generation (2023)

Our recent work also applies language technologies in interdisciplinary settings, including scientific literature analysis, engineering education, and large-scale environmental data.

Research support

We gratefully acknowledge support from: