mma ~ $

# Education

2026 - Present

Ph.D. Computer Science, University of Virginia

Advisor: Prof. Xi Peng

2020 - 2025

Ph.D. Computer Science, University of Delaware

Advisor: Prof. Xi Peng

2019

M.Sc. Electrical Engineering, University of Southern California

2016

B.Sc. Electrical Engineering, Northwest A&F University

# Experience

2026 - Present

Research Assistant, Deep Robust & Explainable AI Lab, University of Virginia

2025 - Present

Visiting Graduate Research Assistant, Memorial Sloan Kettering Cancer Center

2019 - 2025

Research Assistant, Deep Robust & Explainable AI Lab, University of Delaware

# Publications

Medical AI Encodes a "Feeling of Error": Verifying Cancer Segmentation via Internal Concepts

Mengmeng Ma, Yunxiang Peng, Tang Li, Lu Lin, Binsheng Zhao, Oguz Akin, Xi Peng.

ECCV 2026

PDF

Cancer segmentation models reveal a latent “feeling of error”: failed predictions activate fewer and weaker internal concepts. Sparse autoencoders expose these concepts, enabling accurate and interpretable failure detection across three cancer types while preserving segmentation quality.

@inproceedings{ma2026medsae, title={{Medical AI Encodes a "Feeling of Error": Verifying Cancer Segmentation via Internal Concepts}}, author={Mengmeng Ma and Yunxiang Peng and Tang Li and Lin Lu and Volkan Beylergil and Binsheng Zhao and Oguz Akin and Xi Peng}, booktitle={Proceedings of the European Conference on Computer Vision (ECCV)}, year={2026}, }

Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers

Tang Li, Yanlin Chen, Mengmeng Ma, Xi Peng.

ICML 2026

PDF

ViSAE efficiently maps sparse-autoencoder features in Vision Transformers to visual concepts and causally traces how those concepts interact across layers. Its concept circuits reveal spurious reasoning and support steering that improves worst-group accuracy on WaterBirds by 48.2%.

@inproceedings{li2026vase, title = {{Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers}}, author = {Li, Tang and Chen, Yanlin and Ma, Mengmeng and Peng, Xi}, booktitle = {Proceedings of the International Conference on Machine Learning (ICML)}, year = {2026} }

Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings

Yunxiang Peng, Mengmeng Ma, Ziyu Yao, Xi Peng.

CVPR 2026

Highlight PDF

A Vision Transformer's internal causal circuits can serve as label-free proxies for generalization under distribution shift. Dependency Depth Bias improves pre-deployment model selection, while Circuit Shift Score improves post-deployment monitoring, outperforming output-based proxies in both settings.

@inproceedings{peng2026inside, title = {{Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings}}, author = {Peng, Yunxiang and Ma, Mengmeng and Yao, Ziyu and Peng, Xi}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} }

"Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence

Mengmeng Ma, Tang Li, Yunxiang Peng, Lu Lin, Volkan Beylergil, Binsheng Zhao, Oguz Akin, Xi Peng.

ICML 2025

PDF

The paper introduces 180K expert-evaluated image–mask–rationale triples and a rationale-informed objective that learns to disentangle and localize clinical concepts without pixel-level concept labels. The resulting models predict tumors while also explaining why and showing where the supporting evidence lies.

@inproceedings{Ma2025tumor, title = {{``Why Is There a Tumor?'': Tell Me the Reason, Show Me the Evidence}}, author = {Ma, Mengmeng and Li, Tang and Peng, Yunxiang and Lin, Lu and Beylergil, Volkan and Zhao, Binsheng and Akin, Oguz and Peng, Xi}, booktitle = {ICML}, year = {2025} }

Beyond Accuracy: Ensuring Correct Predictions with Correct Rationales

Tang Li, Mengmeng Ma, Xi Peng.

NeurIPS 2024

PDF

Accuracy alone can hide correct predictions made for the wrong reasons. This work curates structured visual rationales and uses rationale-informed optimization to produce “double-correct” models, improving prediction accuracy by up to 10.1% as well as rationale localization and disentanglement.

@inproceedings{Li2024beyond, title = {Beyond Accuracy: Ensuring Correct Predictions with Correct Rationales}, author = {Li, Tang and Ma, Mengmeng and Peng, Xi}, booktitle = {NeurIPS}, year = {2024} }

DEAL: Disentangle and Localize Concept-level Explanations for VLMs

Tang Li, Mengmeng Ma, Xi Peng.

ECCV 2024

Strong Double Blind PDF

DEAL trains Vision-Language Models to make fine-grained concept explanations distinct and correctly localized without human concept annotations. Better concept explanations also reduce reliance on spurious correlations and improve prediction accuracy.

@inproceedings{Li2024deal, title = {DEAL: Disentangle and Localize Concept-level Explanations for VLMs}, author = {Li, Tang and Ma, Mengmeng and Peng, Xi}, booktitle = {ECCV}, year = {2024} }

Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients

Mengmeng Ma, Tang Li, Xi Peng.

ICML 2024

PDF

Topology-aware Federated Learning privately infers relationships among clients, then uses that topology to identify influential clients during optimization. This yields scalable models that generalize more robustly to clients outside the original federation.

@inproceedings{Ma2024federation, title = {Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients}, author = {Ma, Mengmeng and Li, Tang and Peng, Xi}, booktitle = {ICML}, year = {2024} }

Are Data-driven Explanations Robust Against Out-of-distribution Data?

Tang Li, Fengchun Qiao, Mengmeng Ma, Xi Peng.

CVPR 2023

PDF

Models can remain accurate under distribution shift while their explanations become unreliable. Distributionally Robust Explanations uses self-supervised signals across distributions—without human explanation labels—to improve both explanation consistency and prediction robustness on out-of-distribution data.

@inproceedings{Li2023datadriven, title = {Are Data-driven Explanations Robust Against Out-of-distribution Data?}, author = {Li, Tang and Qiao, Fengchun and Ma, Mengmeng and Peng, Xi}, booktitle = {CVPR}, year = {2023} }

Are Multimodal Transformers Robust to Missing Modality?

Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, Xi Peng.

CVPR 2022

PDF

Multimodal Transformers are not inherently robust to missing modalities, and the best fusion strategy varies by dataset rather than universally. The paper therefore automatically searches for a data-specific fusion strategy, improving missing-modality robustness across three benchmarks.

@inproceedings{Ma2022multimodal, title = {Are Multimodal Transformers Robust to Missing Modality?}, author = {Ma, Mengmeng and Ren, Jian and Zhao, Long and Testuggine, Davide and Peng, Xi}, booktitle = {CVPR}, year = {2022} }

SMIL: Multimodal Learning with Severely Missing Modality

Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, Xi Peng.

AAAI 2021

PDF

SMIL uses Bayesian meta-learning to handle severely missing modalities during training, testing, or both—even when 90% of training examples are incomplete. It achieves state-of-the-art results over generative baselines on three multimodal benchmarks.

@inproceedings{Ma2021smil, title = {SMIL: Multimodal Learning with Severely Missing Modality}, author = {Ma, Mengmeng and Ren, Jian and Zhao, Long and Tulyakov, Sergey and Wu, Cathy and Peng, Xi}, booktitle = {AAAI}, year = {2021} }

# Honors & Awards

2026

CVPR’26 Outstanding Reviewer

Top 5% in 17,491 reviewers

2025

Department Travel Award for Outstanding Conference Publications

$2,500 Travel Grant Awarded by Department of Computer & Information Sciences at the University of Delaware

2024

Department Travel Award for Outstanding Conference Publications

$2,500 Travel Grant Awarded by Department of Computer & Information Sciences at the University of Delaware

2022

University Professional Development Award

$3,000 Travel Grant Awarded by University of Delaware Graduate College

2021

Distinguished Graduate Student Award

Department of Computer & Information Sciences at the University of Delaware

# Professional Service

Conference Reviewer

Annual Conference on Neural Information Processing Systems (NeurIPS) 2024, 2025, 2026

International Conference on Machine Learning (ICML) 2025, 2026

International Conference on Learning Representations (ICLR) 2026

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2023, 2024, 2025, 2026

International Conference on Computer Vision (ICCV) 2023, 2025

European Conference on Computer Vision (ECCV) 2024, 2026

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026

International Conference on Artificial Intelligence and Statistics (AISTATS) 2025, 2026

AAAI Conference on Artificial Intelligence (AAAI) 2025, 2026

British Machine Vision Conference (BMVC) 2024, 2025

ACM International Conference on Multimedia (ACM MM) 2020, 2021, 2022, 2023, 2024

Journal Reviewer

Transactions on Machine Learning Research (TMLR)

IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

ACM Computing Surveys (CSUR)

IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)

Neural Networks (NN)

IEEE Transactions on Radiation and Plasma Medical Sciences (TRPMS)

# Teaching

Teaching Assistant, Department of Computer & Information Sciences, University of Delaware

Spring 2022

Introduction to Data Mining (CISC 483/683)

Fall 2021

Introduction to Data Mining (CISC 483/683)

Spring 2021

Introduction to Data Mining (CISC 483/683)

Fall 2020

Introduction to Computer Vision (CISC 442/642), Introduction to Data Mining (CISC 483/683)

# Contact

I'm always interested in hearing about opportunities to collaborate or apply research to real-world problems.

Feel free to reach out.

email wkb9cb [at] virginia [dot] edu

addr 85 Engineer's Way, Rice Hall 436, Charlottesville, VA 22904