Research

With a group at Johns Hopkins, I study how frozen language models handle time: hand a model a hidden clock, and see whether it can act on elapsed time it never reads in the prompt. The work spans residual-stream conditioning, LoRA and FiLM adaptation, causal ablations, and benchmark design.

• ChronoState: hidden elapsed-time conditioning for temporal-state action selection in frozen-backbone language models. First author, arXiv, 2026.
• A second first-author paper on chronometric injection. In preparation.
• Co-author on three more in progress.

Before that, I built Virex, a tweet-virality model trained on 250K+ scraped tweets with a gradient-boosted mixture of experts.