Publications

Research

Workshop papers at NeurIPS 2025 and ICML 2026. Each entry states the question, the method, and the key result; the papers are linked.

NeurIPS 2025 WorkshopSole author

Context-Masked Meta-Prompting for Privacy-Preserving LLM Adaptation in Finance

Question: How can a large language model be adapted to confidential financial tasks without exposing that data to the model?
Method: Iterative meta-prompting with feeder and propagation regeneration, optimizing hard prompts under context masking. Evaluated with GPT-3.5 Turbo on SQuAD, CNN/DailyMail, and SAMSum as proxies for financial Q&A, news, and client-interaction summarization.

Key result

103.87%

improvement in ROUGE-L F1

question answering, GPT-3.5 Turbo

ICML 2026 Workshop

Sophia Abraham, Taiye Chen, Cyril Chhun, Giovanna Jaramillo-Gutierrez, Simon Mylius, Sayash Raaj, Peter Slattery, Sean McGregor

Question: How can the risk of AI incidents be measured when deployment prevalence and reporting rates are unknown?
Method: A public-health surveillance framing of noisy, incomplete incident data, with expert panels combining domain expertise, incident data, and statistical tools. Demonstrated on autonomous-vehicle and deepfake case studies.

Key result

phases of incident emergence

Framework result