Independent researcher measuring frontier LLM/VLM failure modes: context-driven valuation bias, author misattribution, miscalibrated confidence, and confabulation. I design behavioral experiments that find AI failures standard benchmarks miss, and verify every result by hand.
Background: first-author neuroscience research (PNAS 2007, Salk Institute), epidemiology training (MS, UCLA; MPH, Seoul National University), and five years teaching college biology.
Papers
- Which Answer Did the Seventeen-Year-Old Write? Author-age misattribution, miscalibrated confidence, and non-corrective self-explanation in six frontier LLMs (3,225 scored responses). arXiv:2610.31946 · code and data
- Does the Outfit Change the Price? Context halo in vision-language model valuation of an identical low-cost necklace (1,164 sessions, 4,604 logged rows). arXiv:2610.34591 · code and data
In progress
- population-reasoning-shift-llm-pilot: do LLM personas shift how letters are written, and do LLM judges read letters through population stereotypes? (Japanese/American, human/AI; pilot)