Research Interests

I investigate how language models internally represent information and produce decisions, particularly when their outputs are sensitive to irrelevant context, role framing, or subtle variations in input. My research combines behavioral evaluation with mechanistic analysis to identify the internal components and representations associated with model reasoning and failure.

The broader goal is to develop language models that are not only accurate, but also robust, interpretable, and appropriately calibrated for high-stakes decision-making.

Legal and technical documents often contain complex hierarchical structures, dependencies, and domain-specific constraints that conventional language models do not explicitly capture. I develop methods for generating, representing, and evaluating such documents while preserving their structural and semantic coherence.

This research explores structure-aware generation, hierarchical document representation, constrained decoding, and domain-grounded evaluation. The objective is to move beyond surface-level text quality toward models that respect the logical organization, functional relationships, and formal requirements of specialized documents.

Patents provide a challenging environment for AI because they combine technical knowledge, legal reasoning, economic value, and highly structured language. I develop AI methods for patent generation, evaluation, similarity assessment, classification, and valuation.

My research aims to transform complex patent information into transparent and actionable evidence for tasks such as prior-art analysis, claim assessment, technology evaluation, and intellectual-property decision-making. A central priority is ensuring that AI-based patent analysis remains interpretable, legally grounded, and aligned with expert judgment.