I research safety, privacy and alignment in multi-agent LLM systems: benchmark and evaluation design, adversarial probing for privacy leakage, reward modeling, and red-teaming of agentic systems.
full list on Google Scholar ↗A benchmark for what actually drives privacy leakage when LLM agents talk to each other. Across Qwen3, Llama 3.1 and Gemma models, more capable and reasoning models leak more, leakage compounds over turns, and prompt-based mitigations only go so far. Also presented at WiML.
Evaluates whether LLMs respect contextual integrity when one user’s data also reveals something about other people.
An evaluation framework for creativity in LLMs. It surfaces a hive-mind behaviour: different models converge on strikingly similar ideas in creative tasks. Under review at NeurIPS 2026.
Uses causal structure over rubric dimensions to train reward models that are robust to spurious correlations. My dimension-wise rater methodology from Google DeepMind fed into this work.
A dataset and benchmark for tagging numbers in financial filings with XBRL labels, drawn from thousands of possible tags. Built with Goldman Sachs.
Brings structured medical knowledge from the UMLS knowledge graph into natural language inference over clinical text. (* equal contribution)
