A mechanistic-interpretability study testing whether warmth and competence form linearly probeable directions in open-weight LLMs, and whether activation steering along those directions shifts hiring callback decisions.
An end-to-end pipeline that turns English news articles into a fine-tuning-ready dataset for sentence-level media bias detection — RSS crawl, cleaning, a proxy classifier, and LLM+VLM annotation.
Reproducing sentence-level media bias classification on BABE. A clean RoBERTa-base fine-tune that beats published baselines at 0.857 macro-F1.
An audit of the Media Bias Identification Benchmark — label noise probes, near-duplicate detection, and LLM review across 8 tasks.
A study of the rec-dating dataset as a role-based bipartite network, separating outgoing rating activity from received attention and applying HITS consistently.
Full analytical workflow for a paper on rescuing Community Notes stuck in 'Needs More Ratings' through clustering, rescue simulation, topic modeling, and LLM validation.