University of Maryland
academicThe University of Maryland's foundation-model output is dominated by Tom Goldstein's group (tomg-group-umd), director of the UMD Center for Machine Learning: the depth-recurrent latent-reasoning model Huginn-3.5B (co-led with Jonas Geiping's ELLIS/MPI Tübingen group and LLNL — a genuinely multi-lab pretrain), the Gemstones suite of 4,000+ open checkpoints for multi-faceted scaling-law research, the DynaGuard dynamic guardian models (with Capital One), the standard open LLM-text detector Binoculars, and the memorization-mitigating goldfish loss.
Beyond Goldstein: Furong Huang works on trustworthy agents and test-time alignment (GenARM, ICLR 2025); Soheil Feizi's line probes the robustness limits of AI-text detection and watermarking; Jordan Boyd-Graber's QA group seeded the benchmark-science critique of multiple-choice evaluation (ACL 2025); and Schulhoff's UMD-led Prompt Report became the standard practitioner taxonomy of prompting. Hal Daumé III directs the NSF TRAILS institute on trustworthy AI.
People
- Tom Goldstein Google Scholar — Volpi-Cupal Professor; Director, UMD Center for Machine Learning — Huginn, Gemstones, DynaGuard, Binoculars
- Furong Huang Google Scholar — Associate Professor — trustworthy agents / RL, test-time alignment (GenARM)
- Soheil Feizi Google Scholar — Professor — robustness of AI-text detection, watermarking, attribution
- Jordan Boyd-Graber Google Scholar — Professor (CS/UMIACS/iSchool) — QA and benchmark-critique line
- Hal Daumé III Google Scholar — Professor; Director, NSF TRAILS Institute (formerly Microsoft Research NYC (long affiliation; Scholar now lists UMD only))