Make interpretability provable enough to test and actionable enough to use.
PAI Lab advances interpretability through three linked goals: provable foundations, actionable evidence, and tools people can use in real workflows.
Learn more about PAIProvable. Actionable. Interpretability.
PAI Lab
Provable methods. Actionable evidence. Interpretability tools for real AI systems.
PAI turns explanations into testable evidence and usable interventions.
Provable
Explanations grounded in assumptions, guarantees, and visible limits.
Actionable
Evidence that helps people debug, decide, intervene, and discover.
Interpretability
Mechanisms, evidence, and interfaces that make model behavior legible in real workflows.
Research
View all research Five pillars, one PAI agenda.
Theory, large models, systems, interaction, and science connect provable foundations with actionable interpretability.
01 Foundations
Theoretical Foundations of Usable XAI
When is an explanation provable, useful, and theoretically justified?
02 Large ModelsUseful XAI in Large Models
How can we understand and intervene in language, vision, and multimodal models?
03 SystemsSystems for XAI
What infrastructure makes interpretability scalable, reproducible, and deployable?
04 InteractionInteractive XAI
How should humans query, evaluate, refine, and act on explanations?
05 ScienceXAI for Science
How can interpretability support scientific discovery and hypothesis generation?
Selected Papers
View all publications Recent work across provable, actionable interpretability.
Representative papers from each part of the PAI agenda.
Theory
Bayesian Gated Non-Negative Contrastive Learning
ICML 2026Builds gated Bayesian contrastive objectives for more reliable representation-level evidence.
Large ModelsBeyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability
ICML 2026Evaluates LLM reasoning through geometric progress and stability rather than scalar-only scores.
SystemsTowards Reasoning-Preserving Unlearning in Multimodal Large Language Models
CVPR 2026Studies unlearning systems that remove target data while preserving multimodal reasoning behavior.
Interactive XAIAlgorithmic Recourse of In-Context Learning for Tabular Data
ICML 2026Connects in-context learning with actionable recourse for tabular decision support.
ScienceControlling Repetition in Protein Language Models
ICLR 2026Controls repetitive generation behavior in protein language models for more useful scientific design.
- Bayesian Gated Non-Negative Contrastive Learning. ICML 2026 Theory
- Understanding Generalization and Forgetting in In-Context Continual Learning. ICML 2026 Theory
- Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability. ICML 2026 Large Models
- 2026-05 One paper accepted at TMLR
- 2026-04 Eight papers accepted at ICML 2026
- 2026-04 Three papers accepted at ACL 2026
Work on provable, actionable interpretability.
We are looking for curious people who want to connect foundations, systems, and real-world use.
Open Positions