Yi Ding

Yi Ding 丁熠

PhD Student · Computer Science · Purdue University

— DripNowhy

Yi Ding (丁熠)

About

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

Research Interests

  • LLM/VLM Post-Training

    Reinforcement Learning, On-Policy Distillation, Reasoning, Self-Correction, and Self-Evolving.

  • Alignment and Reliability

    Preference Alignment, Safety Alignment, Reliable Supervision, and Robustness.

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

News

  1. Aug 2026

    Our paper on on-policy self-adaptation (OPSA) was released.

  2. May 2026

    Two papers were accepted to ICML 2026.

  3. Jan 2026

    One paper was accepted by ICLR 2026.

  4. Sep 2025

    One paper was accepted by NeurIPS 2025.

  5. Aug 2025

    One paper was accepted by EMNLP 2025 Main Conference.

  6. May 2025

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

  7. Apr 2025

    Yi serves as reviewer for NeurIPS 2025.

  8. Jan 2025

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

  9. Jan 2025

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

  10. Sep 2024

    Yi serves as reviewer for ICLR 2025.

  11. Sep 2024

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

  12. Jul 2024

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

  13. May 2024

    Our paper about multimodal fusion was accepted at ICML 2024.

Publications * equal contribution

11 papers

Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control
Preprint 2026

Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control

Bolian Li, Yifan Wang, Yi Ding, Anamika Lochab, Ananth Grama, Ruqi Zhang

We propose Entrocraft, a rejection-sampling method that precisely controls entropy schedules during LLM reinforcement learning, alleviating performance saturation while improving generalization, output diversity, and long-term training.

Paper Code Demo
Learning Self-Correction in Vision–Language Models via Rollout Augmentation
ICML 2026

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

Yi Ding, Ziliang Qiu, Bolian Li, Ruqi Zhang

Learning self-correction by augmenting reinforcement-learning rollouts.

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

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

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.

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

Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models

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

Teaching vision-language models to reason about safety across multiple images.

Paper Code Project
SafeWork-R1: Coevolving Safety and Intelligence under the AI-45° Law
Technical Report 2025

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

Shanghai Artificial Intelligence Laboratory, …, Yi Ding, and 100+ authors

Jointly improving safety and intelligence in multimodal reasoning.

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

Sherlock: Self-Correcting Reasoning in Vision-Language Models

Yi Ding, Ruqi Zhang

Bootstrapping self-correction and self-improvement from limited supervision.

Paper Code Project
Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection
EMNLP 2025 · Main

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

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

We present VisCo-Attack, which jailbreaks MLLMs via a visual-centric setting and fabricated visual context.

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

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

Yi Ding, Bolian Li, Ruqi Zhang

Training-free safety alignment for vision-language models at inference time.

Paper Code Project
Test-Time Dynamic Image Fusion
NeurIPS 2024

Test-Time Dynamic Image Fusion

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

Improving the quality of fused images across almost every backbone without additional training, by setting dynamic weights at test time.

Paper Code
Predictive Dynamic Fusion
ICML 2024

Predictive Dynamic Fusion

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

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

Paper Code

GitHub

Loading repositories…

Education

2025.08 — Present

Ph.D. in Computer Science, Purdue University

Advisor: Dr. Ruqi Zhang

2021.08 — 2025.06

B.S. in Mathematics, Tianjin University

Advisor: Dr. Bing Cao

Academic Services

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