Along the way
News.
New papers, research milestones, and things I've been working on.
-
Publication
Sycophancy Suppression Can Impair Rational Updating is now on arXiv, accepted to EMNLP 2026 Findings. We study how to reduce sycophancy while preserving a model’s ability to learn from useful feedback.
Read more about this August 2026 update -
Preprint
Introducing MemAudit, our benchmark for measuring what AI agents actually remember about their users across interactions.
Read more about this June 2026 update -
Survey
Our survey, The Landscape of Agentic Time Series Systems, is available, covering system architectures, reliability, and open research directions.
Read more about this June 2026 update -
Preprint
ARC-STAR is on arXiv. We develop auditable corrections for PDE foundation models, focusing computation on regions that need refinement.
Read more about this May 2026 update -
Publication
Locally Confident, Globally Stuck is on arXiv, studying quality versus exploration in diffusion language models. Update: accepted to COLM 2026 in July.
Read more about this April 2026 update -
Preprint
CellMaster is now on arXiv: an AI agent for interpretable cell type annotation, with support for collaborative human refinement.
Read more about this February 2026 update -
Preprint
The original PAN preprint is released: a world model for interactive, long-horizon simulation guided by natural language actions.
Read more about this November 2025 update -
Publication
scPilot is a NeurIPS 2025 poster. Our framework brings LLM reasoning to single-cell analysis, supported by the scBench evaluation suite.
Read more about this September 2025 update