Pang Wei Koh
Assistant Professor
Develops reliable machine learning methods, including robustness to distribution shift and retrieval-based language models.
Pang Wei Koh's lab at UW works on trustworthy & safe ML, large language models and knowledge, QA & retrieval. 62 papers since 2023, 38 at top venues, mostly NeurIPS and ICML. 8 PhD students. 11 alumni, most often to Anthropic and Meta. Applications close .
At a glance
- papers, 2023–2026
- 16/ year
- at T1 venues
- 61%
- PhD students
- 8
- MS students
- 0
- alumni listed
- 11
Papers by yearfrom the lab homepage, 2023–2026
- 18 papers2026 · T1 11 · collecting
- 21 papers2025 · T1 12
- 16 papers2024 · T1 10
- 7 papers2023 · T1 5
T1 venuesOtherVenues: NeurIPS, ICML, ICLR, EMNLP, ACL
Research focuspaper titles that use each topic's words
- Large language models25 papers
- Knowledge, QA & retrieval4 papers
- Trustworthy & safe ML1 paper
- shift0 papers
- distribution shift0 papers
Paper list
From the lab homepage.
202618 papers
- SatMLA false sense of privacy: Evaluating textual data sanitization beyond surface-level privacy leakage (opens in a new tab)
- CHILALMo: Interactive Aim-Limit-defined, Multi-Objective system for personalized high-dose-rate brachytherapy treatment planning and visualization for cervical cancer (opens in a new tab)
- ICML (T1)Anchored decoding: Provably reducing copyright risk for any language model (opens in a new tab)
- ACM Technology Policy Council TechBriefBuy versus build an LLM: A decision framework for governments (opens in a new tab)
- ICML (T1)Cold-start personalization via training-free priors from structured world models (opens in a new tab)
- NeurIPS (T1)Deep reasoning in general purpose agents via structured meta-cognition (opens in a new tab)
- ICML (T1)DR Tulu: Reinforcement learning with evolving rubrics for deep research (opens in a new tab)
- Journal of Clinical OncologyEnabling federated discovery leveraging agentic AI systems: The Cancer AI Alliance and DataVoyager. (opens in a new tab)
- ICML WorkshopFabScore: Fine-Grained Evaluation of Fabrications in Automated AI Research (opens in a new tab)
- ICLR (T1)Frustratingly simple retrieval improves challenging, reasoning-intensive benchmarks (opens in a new tab)
- NeurIPS (Evaluations and Datasets Track)JobBench: Aligning agent work with human will (opens in a new tab)
- Microsoft AIMAI-Thinking-1: Building a Hill-Climbing Machine (opens in a new tab)
202521 papers
- COLM2 OLMo 2 Furious (opens in a new tab)
- ICML (T1)DataDecide: How to predict best pretraining data with small experiments (opens in a new tab)
- COLMEstablishing task scaling laws via compute-efficient model ladders (opens in a new tab)
- COLMEvalTree: Profiling language model weaknesses via hierarchical capability trees (opens in a new tab)
- ACL (T1)Exploring how generative MLLMs perceive more than CLIP with the same vision encoder (opens in a new tab)
- NeurIPS (T1)FlexOlmo: Open language models for flexible data use (opens in a new tab)
- COLMFluid language model benchmarking (opens in a new tab)
- ICLR (T1)Group-robust sample reweighting for subpopulation shifts via influence functions (opens in a new tab)
- NeurIPS WorkshopHybridCoT: Interleaving latent and text chain-of-thought for efficient reasoning (opens in a new tab)
- ICLR (T1)Language models scale reliably with over-training and on downstream tasks (opens in a new tab)
- ICML (T1)NICE: Non-differentiable evaluation metric-based data selection for instruction tuning (opens in a new tab)
- ICLR (T1)OLMoE: Open Mixture-of-Experts language models (opens in a new tab)
202416 papers
- EMNLP (T1)Annotation alignment: Comparing LLM and human annotations of conversational safety (opens in a new tab)
- EMNLP (T1)CopyBench: Measuring literal and non-literal reproduction of copyright-protected text in language model generation (opens in a new tab)
- Findings of EMNLPData-centric AI in the age of large language models (opens in a new tab)
- NeurIPS (Datasets and Benchmarks Track)DataComp-LM: In search of the next generation of training sets for language models (opens in a new tab)
- Proceedings of the National Academy of Sciences (PNAS)Impossibility theorems for feature attribution (opens in a new tab)
- COLMInformation-theoretic distillation for reference-less summarization (opens in a new tab)
- NAACL (T1)Instructional fingerprinting of large language models (opens in a new tab)
- ICLR (T1)Leveraging domain relations for domain generalization (opens in a new tab)
- NeurIPS (T1)MEDIQ: Question-asking LLMs for adaptive and reliable clinical reasoning (opens in a new tab)
- Findings of EMNLPMerge to Learn: Efficiently adding skills to language models with model merging (opens in a new tab)
- NeurIPS (T1)Multilingual diversity improves vision-language representations (opens in a new tab)
- NeurIPS (T1)Scaling retrieval-based language models with a trillion-token datastore (opens in a new tab)
20237 papers
- NeurIPS (T1)Are aligned neural networks adversarially aligned? (opens in a new tab)
- NeurIPS (Datasets and Benchmarks Track)DataComp: In search of the next generation of multimodal datasets (opens in a new tab)
- EMNLP (T1)FActScore: Fine-grained atomic evaluation of factual precision in long form text generation (opens in a new tab)
- NeurIPS (T1)On the trade-off of intra-/inter-class diversity for supervised pre-training (opens in a new tab)
- ICML (T1)Out-of-distribution robustness via targeted augmentations (opens in a new tab)
- NeurIPS (T1)Proximity-informed calibration for deep neural networks (opens in a new tab)
- NeurIPS WorkshopRetrieval-based language models using a multi-domain datastore (opens in a new tab)
Where alumni went11 listed, 43% to industry
Industry 3Academia 4
Most often to
- Anthropic
- Meta
Also to
- Stanford University
- NUS
- Princeton University
Apply to UWFall 2027
UW Allen School Computer Science PhD
- Round
- Single round
- Opens
- GRE
- not required
- English
- See program page
- Fee
- See program page
Not yet confirmed for this intake.
Similar labsnearest by research summary
- SGNUS
Tat-Seng Chua
- Large language models
- GBEdinburgh
Wenda Li
- Reasoning & neuro-symbolic AI
- USUCSD
Prithviraj Ammanabrolu
- Reasoning & agents
- USUCLA
Vwani Roychowdhury
- Language & society
- CNTsinghua
Qingyao Ai
- Large language models
- USMichigan
Lu Wang
- Interpretability, evaluation & safety of LMs
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
UWcloses