Kat Dearstyne

Software Engineering & AI Researcher

PhD student at Notre Dame researching how AI can improve software engineering and how better software practices can make AI safer. Passionate about AI safety and translating research into practical, usable tools.

Learn More Download CV
Kat Dearstyne

Research Projects

Encoded Reasoning in Chain-of-Thought

Current

Investigating how constraints on chain-of-thought affect LLM performance and incentivize encoded reasoning, with implications for the monitorability of reasoning in modern models.

Requirements Engineering for Code Generation

Current

Empirically investigating the application of requirements engineering practices to code generation, examining how to increase maintainability of generated code and reduce common bug patterns.

Faithfulness of Cross-Layer Transcoders

Past

Investigated memorization and generalization in LLMs using mechanistic interpretability as part of the SPAR fellowship under Georg Lange. Identified a limitation in Cross-Layer Transcoders, where they can be incentivized to learn unfaithful circuits.

View Post

More Traceable Software Agents for Human Oversight

Past

Collaborated with Microsoft Research to enhance human oversight of AI agents in software engineering tasks by providing traceability to key project artifacts that informed decision-making.

View Publication

Designing More Trustworthy RL Systems

Past

Created a framework for designing more trustworthy reinforcement learning systems and stronger system requirements as part of NSF's Designing Accountable Software Systems (DASS) initiative.

View Publication

Automated Traceability & Requirements Analysis

Past

Enhanced automation of requirements analysis, trace link discovery, and documentation generation. Contributed to building a visualization platform for commercialization through NASA STTR Phase I and the Techstars Accelerator Program.

View Publications

Recent Publications & Writing

2026

Cross-Layer Transcoders are Incentivized to Learn Unfaithful Circuits

LessWrong

2026

Revealing the Dark Matter: Connecting Tacit and System Knowledge in Human-AI Collaborations

ACM/IEEE International Conference on Software Engineering (ICSE), NIER Track

2026

Real-World Traceability Patterns for Generative AI Systems

RAISE Workshop on Requirements Engineering for AI-powered SoftwarE at ICSE

2025

QuestRL: A Q&A Framework for Designing Trustworthy Reinforcement Learning Systems

IEEE 33rd International Requirements Engineering Conference

View All Publications