This page is a public record of projects and collaborations.
Medmarks
Contributed to an open-source benchmark suite for medical LLM capabilities.
Work
Projects and collaborations.
Selected work
AI Safety Research Collaborations
Research work on self-preservation propensity, emergent misalignment, and character training.
Adaptive Sampling Networks
Learning distribution-dependent decoding rules for frozen language models.
Normative Drift in Emergent Misalignment
Narrow fine-tuning that induces emergent misalignment also shifts how models judge alignment — often by collapsing discrimination rather than by endorsing harm.