Research
I study when and why language models fail and how to fix them, with a focus on agents that are trustworthy, secure, and safe.
I identify real problems that undermine trust in AI systems, build evaluations to find when and why they occur, and use those findings to design targeted mitigations. Most of my current work is on secure and reliable agents.
Finding when and why models fail
- SecFid separates a model handling an injected instruction as data from a model silently dropping it. Standard prompt-injection benchmarks count both as secure.
- That’s Deprecated! studies what code models do when a prompt contradicts what they learned, and whether that conflict is visible in their activations. I led the dataset construction and evaluation.
Targeted mitigations
- Fidelity-aware preference training uses SecFid’s scoring as a training signal, teaching a defended model to keep injected text as data instead of dropping it.
- Activation steering shifts whether a code model follows the prompt or its prior knowledge. I led this part of the That’s Deprecated! project.
Papers
- 2026
- 2025
That’s Deprecated! Understanding, Detecting, and Steering Knowledge Conflicts in Language Models for Code Generation
My role: led dataset construction, evaluation, and the steering mitigation.
- 2022
Progressive Alignment of Crystals: Reproducible and Efficient Assessment of Crystal Structure Similarity
My role: collected and analyzed PAC results.
Open source
-
Unroll is a free VS Code extension for reading agent traces and eval runs. It shows each run as a conversation and links every step to the JSON record it came from.