Foad Namjoo
Computer Science PhD student since January 2023 at the Kahlert School of Computing, University of Utah, advised by Professor Jeff Phillips in the UtahDB Lab. I work on LLM interpretability and activation steering, high-dimensional statistics, and wearable health sensing.
Selected work
LLM benchmark audit and cleaned dataset
Answer style alone: 68.9% on TruthfulQA, where chance is 50%
First-author paper PyTorch Transformers scikit-learnLLM steering datasets
3 public datasets on Hugging Face, used in 2 papers
Creator Hugging Face DatasetsMotionPI wearable health platform
Heart, motion, surveys and GPS: 7.7M records a day
Sole developer of the app, server, and sync first-author paper Flutter BLE Node.js MongoDBdKS distribution test library
~76,000× faster than exact computation
Co-authored paper open source C++ pybind11Spatial hotspot detection
All contiguous U.S. counties scanned in 0.33 s
First-author paper ACM SIGSPATIAL 2026 pyscanNews
- Accepted Understanding and Mitigating Dataset Corruption in LLM Steering, led by Cullen Anderson, was accepted to Findings of EACL 2027. Nov 13, 2026
- Presentation Presented Sampling for Region-Aggregated Spatial Scan Statistics at ACM SIGSPATIAL 2026 in Riverside, CA. Slides. Nov 5, 2026
- Visit Visited Esri headquarters in Redlands, CA, the company behind ArcGIS, during ACM SIGSPATIAL 2026, to explore how AI agents and GeoAI are changing spatial analysis, and why human judgment still matters. Nov 3, 2026
- Preprint New on arXiv: Orthogonal and Parametrized Multi-Attribute Steering of LLMs, led by Remy Ogasawara — clipping each steering vector to its most steerable range, setting a trait's level rather than pushing it, and orthogonalizing correlated directions, so several traits steer at once without over- or under-steering. Code. Oct 27, 2026
- Preprint New on arXiv: Judging by the Cover — a six-feature surface probe that never reads the question separates TruthfulQA's correct from incorrect answers at AUC 0.715; Audit-Prune removes the leakage and we release TruthfulQA-476. Sep 2026
- Datasets Released TruthfulQA-476 on Hugging Face — a surface-cleaned subset of binary-choice TruthfulQA (476 of 790 pairs, Apache-2.0), with two held-out evaluation sets, SurfaceFlipped-131 and Natural-131. Sep 2026
- Published The MotionPI system paper is out in Springer LNICST vol. 699, pp. 283–293 — the secure, resilient design behind an NIH-funded field study collecting wearable and survey data in the wild. Sep 2, 2026
- Accepted Sampling for Region-Aggregated Spatial Scan Statistics was accepted as a full paper at ACM SIGSPATIAL 2026 — 58 of 248 submissions. Code. Aug 2026
- Preprint New on arXiv: Sampling for Region-Aggregated Spatial Scan Statistics — replacing each region with 20–50 sampled points restores the statistical power that centroid-collapsing discards. Code. Jul 1, 2026
- Release Released dKS — a fast, header-only C++ library with Python bindings (
import dks) for the multi-dimensional Kolmogorov–Smirnov distance, with both exact and near-linear-time algorithms. Paper · Code. Jun 15, 2026 - Visit Visited Martian AI, a startup building LLM model routers, for a research day of talks on interpretability, steering, and evaluation. Jan 16, 2026
- Presentation Presented the MotionPI paper at the SmartSP 2025 conference. Dec 1, 2025
- Datasets Released three contrast datasets on Hugging Face. Nov 12, 2025
- Poster Presented dKS research at the Utah AI Summit 2025. Jun 18, 2025