Wenbo Hu's Homepage (胡文波)
 |
Email: hwb13 [AT] tsinghua [DOT] org [DOT] cn |
About
Associate Professor, School of Computer Science and Information Engineering, Hefei University of Technology
My research lies in the statistical machine learning. For perspectives, I focus on trustworthy machine learning, incluing Probabilistic Uncertainty, adversial robustness, causality, etc. For applications, I am interested in time series analysis, computer vision, natural language processing, etc.
[Selected Papers] [Selected Talks & Tutorials]
News!
[Tutorial] I will give a tutorial on Uncertainty and Calibration in Large Models – Reliable Reasoning and Agent Safety at CNCC 2026 on October 21, 2026. (Oct/26)
[Paper] Our recent paper on Visual EQA-CL: A Continual Learning Task for Embodied Question Answering is accepted to Multimedia Systems. (Sep/26)
[preprint] Our recent preprint on Resolving sources of uncertainty in AI weather forecasting has been substantially revised and updated, with model and code made public. (Sep/26)
[Paper] Our recent paper on Visual Token Compression Enhances Robustness of MLLMs is accepted to ACMMM-2026. See you in Brazil! (July/26)
[Talk] I gave a talk at Huawei Noah's Ark Lab with the topic Trustworthiness, Robustness, and Evaluation of Multimodal Large Language Models (Feb/26)
[Paper] Our recent paper on ProSafePrune: Projected Safety Pruning for Mitigating Over-Refusal in LLMs is accepted to ICLR-2026. See you in Brazil! (Jan/26)
[preprint] Our recent preprint Bridging the Gap Between Bayesian Deep Learning and Ensemble Weather Forecasts is online. (Nov/25)
[Paper] Three papers are accepted to AAAI-2026. See you in Singapore! (Nov/25)
Selected Papers
Unveiling Uncertainty: A Deep Dive into Calibration and Performance of Multimodal Large Language Models – COLING 2025.
We investigate how well multimodal models assess their own uncertainty and improve calibration through temperature scaling and prompt optimization.
[Paper] [Code]
Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning – AAAI 2026.
We use truthfulness-sensitive internal activations to estimate step-level confidence and guide more reliable chain-of-thought search.
[Paper] [Code]
ProSafePrune: Projected Safety Pruning for Mitigating Over-Refusal in LLMs – ICLR 2026.
We reduce over-refusal through subspace-projected low-rank pruning while preserving refusal of genuinely harmful requests.
[Paper] [Code]
Visual Token Compression Enhances Robustness of MLLMs – ACM MM 2026.
We prune misaligned visual tokens to improve robustness against jailbreaks and hallucinations while reducing inference cost.
[Paper] [Code]
Resolving sources of uncertainty in AI weather forecasting (Pangu-Bayes) – Preprint, updated September 2026.
We model atmospheric-state and learned-model uncertainty separately to study their different roles in tropical cyclone track and intensity prediction.
[Paper] [Code] [Model]
Selected Talks & Tutorials
Uncertainty and Calibration in Large Models – Reliable Reasoning and Agent Safety
CNCC 2026 Tutorial, October 21, 2026 (upcoming).
An introduction to uncertainty quantification and calibration, with applications to reliable reasoning and agent safety.
[Slides] [Program]
Trustworthiness, Robustness, and Evaluation of Multimodal Large Language Models
Huawei Noah's Ark Lab, February 2026.
A research talk on the reliability, adversarial robustness, and evaluation of multimodal large language models.
[Slides]