About Me

Hi there! I'm a PhD student at Purdue University Purdue Logo, Department of Computer Science, advised by Dr. Ruqi Zhang. I obtained my B.S. degree at the School of Mathematics, Tianjin University TJU Logo. Previously, I worked as a research assistant in the MLDM Lab's Multimodal Vision Processing (MVP) Group, under the guidance of Dr. Bing Cao.

My research interests lie in developing reliable machine learning algorithms and frameworks for real-world applications, with a particular focus on the alignment of Large Foundation Models (LLMs and VLMs) and the generalization of multimodal learning algorithms.

Research Interests

  • Multimodal Learning

    Multimodal fusion and imbalanced multimodal learning.

  • Foundation Models

    Alignment, reasoning, and self-correction for LLMs and VLMs.

  • Trustworthy AI

    Safety, uncertainty, and reliability in real-world model behavior.

Open to collaborations! Feel free to reach out if our research interests align.

Latest News

May 2026

Two papers were accepted to ICML 2026.

Jan 2026

One paper was accepted by ICLR 2026.

Sep 2025

One paper was accepted by NeurIPS 2025.

Aug 2025

One paper was accepted by EMNLP 2025 Main Conference.

May 2025

Yi will give a talk about VLM safety at Shenlan School.

Apr 2025

Yi serves as reviewer for NeurIPS 2025.

Jan 2025

Our paper, dataset, and models about VLM Multi-Image Safety (MIS) are released.

Jan 2025

Our paper about MLLM safety alignment was accepted at ICLR 2025.

Sep 2024

Yi serves as reviewer for ICLR 2025.

Sep 2024

Our paper about dynamic image fusion without additional training was accepted at NeurIPS 2024.

Jul 2024

Yi will present a poster at ICML 2024, Hall C 4-9 #2817, Vienna, Austria.

May 2024

Our paper about multimodal fusion was accepted at ICML 2024.

Publications

* indicates equal contribution.
ICML 2026
Learning Self-Correction in Vision–Language Models via Rollout Augmentation

Learning Self-Correction in Vision–Language Models via Rollout Augmentation

Yi Ding, Ziliang Qiu, Bolian Li, Ruqi Zhang

TL;DR: We propose Octopus, an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. Octopus-8B achieves SoTA performance by advancing reasoning and self-correction capabilities.

International Conference on Machine Learning (ICML), 2026

ICML 2026 Position Paper
Modular Safety Guardrails Are Necessary for Foundation-Model-Enabled Robots in the Real World

Modular Safety Guardrails Are Necessary for Foundation-Model-Enabled Robots in the Real World

Joonkyung Kim, Wenxi Chen, Davood Soleymanzadeh, Yi Ding, Xiangbo Gao, Zhengzhong Tu, Ruqi Zhang, Fan Fei, Sushant Veer, Yiwei Lyu, Minghui Zheng, Yan Gu

TL;DR: We propose modular safety guardrails with monitoring and intervention layers, and show how cross-layer co-design enables faster, less conservative, and more effective safety for physical AI.

ICML 2026 Position Paper

Technical Report
SafeWork-R1: Coevolving Safety and Intelligence under the AI-45◦ Law

SafeWork-R1: Coevolving Safety and Intelligence under the AI-45◦ Law

Shanghai Artificial Intelligence Laboratory, ..., Yi Ding, [and 100+ authors]

TL;DR: We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety.

Technical Report, 2025

ICLR 2026
Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models

Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models

Yi Ding*, Lijun Li*, Bing Cao, Jing Shao

TL;DR: Introducing the first multi-image safety (MIS) dataset, which includes both training and test splits. The VLMs fine-tuned with the MIRage method and MIS training set to improve both the safety and general performance of the models.

International Conference on Learning Representations (ICLR), 2026

NeurIPS 2025
Sherlock: Self-Correcting Reasoning in Vision-Language Models

Sherlock: Self-Correcting Reasoning in Vision-Language Models

Yi Ding, Ruqi Zhang

TL;DR: We present Sherlock, a self-correction and self-improvement training framework enhancing VLM reasoning ability using minimal annotated data.

Neural Information Processing Systems (NeurIPS), 2025

EMNLP 2025
Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection

Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection

Ziqi Miao*, Yi Ding*, Lijun Li, Jing Shao

TL;DR: We present VisCo-Attack, which jailbreak MLLMs via visual-centric setting and fabricated visual context.

EMNLP 2025 Main

ICLR 2025
ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time

ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time

Yi Ding, Bolian Li, Ruqi Zhang

TL;DR: Establishing multimodal safety mechanism for VLMs and enhancing harmlessness and helpfulness of responses without additional training.

International Conference on Learning Representations (ICLR), 2025

NeurIPS 2024
Test-Time Dynamic Image Fusion

Test-Time Dynamic Image Fusion

Bing Cao, Yinan Xia*, Yi Ding*, Changqing Zhang, Qinghua Hu

TL;DR: Improving quality of fused images of almost every backbones without additional training via setting dynamic weight in test-time.

Neural Information Processing Systems (NeurIPS), 2024

ICML 2024
Predictive Dynamic Fusion

Predictive Dynamic Fusion

Bing Cao, Yinan Xia*, Yi Ding*, Changqing Zhang, Qinghua Hu

TL;DR: The key to dynamic fusion lies in the correlation between the weights and the loss, providing generalization theory for decision-level fusion.

International Conference on Machine Learning (ICML), 2024

GitHub Repositories

Education

2025.08 - Present

Ph.D. at Computer Science, Purdue University

Advisor: Dr. Ruqi Zhang

2021.08 - 2025.06

B.S. at School of Mathematics, Tianjin University

Advisor: Dr. Bing Cao

Academic Services

Conference Reviewer
ICLR 2025, 2026 NeurIPS 2025, 2026 ICML 2026 ARR 2025