Projects
Topic: LLM and Foundation Models
One-Shot Safety Alignment for Large Language Models via Optimal Dualization
TL;DR: We propose a one-shot safety alignment algorithm for LLM safety and helpfulness alignment.
Venue Spotlight at Neurips, 2024
[Full Paper] [Code] [Poster]
Moment: A family of open time-series foundation models
TL;DR: We collect a large and diverse collection of time series datasets, and introduce a family of time series foundation models.
Venue: Accepted to ICML, 2024
[Full Paper][Code]
REDO: Execution-Free Runtime Error Detection for COding Agents
TL;DR: We integrate static analysis tools with LLM to detect python runtime errors for LLM-based coding agents.
Venue: In Submission
Topic: Uncertainty Quantification
Uncertainty in Language Models: Assessment through Rank-Calibration
TL;DR: We propose a novel LLM uncertainty quantification metric from the perspective of monotonicity.
Venue: Accepted to EMNLP, 2024
[Full Paper][Code][Poster]
TRAQ: Trustworthy Retrieval Augmented Question Answering via Conformal Prediction
TL;DR: TRAQ utilizes Conformal prediction and Bayesian optimization to guarantee the correctness of RAG. Venue: Accepted at NAACL, 2024
[Full Paper][Code][Poster]
Conformalized credal regions for classification with ambiguous ground truth
TL;DR: TRAQ utilizes Conformal prediction and Bayesian optimization to guarantee the correctness of RAG. Venue: In Submission
Conformal Structured Prediction
TL;DR: This paper introduces a general framework for conformal prediction in structured settings, enabling interpretable prediction sets for complex outputs like text generation and hierarchical labels, while ensuring desired coverage guarantees.
Venue: In Submission
PAC confidence predictions for deep neural network classifiers
TL;DR: We propose a novel algorithm using Clopper-Pearson confidence intervals and histogram binning to construct provably correct classification confidences for deep neural networks, enabling rigorous downstream guarantees in fast inference and safe planning tasks.
Venue: Accepted to ICLR, 2021
Topic: Trustworthy ML
PAC-Wrap: Semi-Supervised PAC Anomaly Detection
TL;DR: We guarantee the false positive and false negative rates of anomaly detection algorithms via conformal prediction.
Venue: Accepted to KDD, 2022
[Full Paper][Code]
Angelic patches for improving third-party object detector performance
TL;DR: This work proposes angelic patches, generated via a reversed FGSM, to significantly enhance object detection robustness, achieving transferable improvements in classification and bounding box accuracy across models and transformations, with a 30% accuracy boost in real-world settings.
Venue: Accepted to CVPR, 2023
Robust model predictive shielding for safe reinforcement learning with stochastic dynamics
TL;DR: We propose a framework for safe reinforcement learning in stochastic nonlinear dynamical systems by integrating a tube-based robust nonlinear model predictive controller as a backup to ensure high-probability safety guarantees, demonstrated on tasks like cart-pole and obstacle navigation.
Venue: Accepted to ICLR, 2020