Jiahao Yu's Page

selected publications

1. arXiv

   Building Coding Agents via Entropy-Enhanced Multi-Turn Preference Optimization

   1st on SWEBench-Lite(Open-weight)
   5th on SWEBench-Verified(Open-weight)

   Jiahao Yu*, Zelei Cheng*, Xian Wu, and 1 more author

   arXiv preprint arXiv:2509.12434 2026

   PDF
2. NIPS

   GPO: Learning from Critical Steps to Improve LLM Reasoning

   Featured in MIT Technology Review China

   Jiahao Yu*, Zelei Cheng, Xian Wu, and 1 more author

   In 2025

   PDF
3. NIPS

   BlockScan: Detecting Anomalies in Blockchain Transactions

   Jiahao Yu*, Xian Wu*, Hao Liu, and 2 more authors

   In 2025

   PDF
4. USENIX

   Mind the Inconspicuous: Revealing the Hidden Weakness in Aligned LLMs’ Ethical Boundaries

   Long Talk

   Jiahao Yu*, Haozheng Luo*, Jerry Yao-Chieh, and 3 more authors

   In Proceedings of the 2025 USENIX Security 2025

   PDF
5. USENIX

   PATCHAGENT: A Practical Program Repair Agent Mimicking Human Expertise

   Long Talk
   Patched over 10 real-world bugs
   CSAW 2025 Finalist

   Zheng Yu, Ziyi Guo, Yuhang Wu, and 5 more authors

   In Proceedings of the 2025 USENIX Security 2025
6. ICML

   The Illusion of Role Separation: Hidden Shortcuts in LLM Role Learning (and How to Fix Them)

   Zihao Wang, Yibo Jiang, Jiahao Yu, and 1 more author

   In Proceedings of the 42nd International Conference on Machine Learning 2025
7. USENIX

   LLM-Fuzzer: Scaling Assessment of Large Language Model Jailbreaks

   Jiahao Yu, Xingwei Lin, Zheng Yu, and 1 more author

   In Proceedings of the 2024 USENIX Security 2024
8. NIPS

   Soft-Label Integration for Robust Toxicity Classification

   Featured in MIT Technology Review China

   Zelei Cheng, Xian Wu, Jiahao Yu, and 3 more authors

   In Proceedings of the 38th Conference on Neural Information Processing Systems 2024

   PDF
9. ICML

   RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation

   Spotlight Top-3.5%

   Zelei Cheng, Xian Wu, Jiahao Yu, and 3 more authors

   In Proceedings of the 41st International Conference on Machine Learning 2024
10. ICLR@SET-LLM

    Assessing Prompt Injection Risks in 200+ Custom GPTs

    Featured in WIRED

    Jiahao Yu, Yuhang Wu, Dong Shu, and 3 more authors

    In ICLR 2024 Workshop on Secure and Trustworthy Large Language Models 2024

    PDF
11. arXiv

    GPTFuzzer: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts

    Geekcon 2023 Annual Themed Debate Breakthrough Awards

    Jiahao Yu, Xingwei Lin, Zheng Yu, and 1 more author

    In 2023

    PDF
12. NIPS

    StateMask: Explaining Deep Reinforcement Learning through State Mask

    Zelei Cheng*, Xian Wu*, Jiahao Yu*, and 3 more authors

    In Proceedings of the 37th Conference on Neural Information Processing Systems 2023

    Bib PDF

    @inproceedings{statemask,
      title = {StateMask: Explaining Deep Reinforcement Learning through State Mask},
      author = {Cheng*, Zelei and Wu*, Xian and Yu*, Jiahao and Sun, Wenhai and Guo, Wenbo and Xing, Xinyu},
      booktitle = {Proceedings of the 37th Conference on Neural Information Processing Systems},
      year = {2023},
    }

13. USENIX

    AIRS Explanation for Deep Reinforcement Learning based Security Applications

    Jiahao Yu, Wenbo Guo, Qi Qin, and 3 more authors

    In Proceedings of the 2023 USENIX Security 2022

    Abs PDF

    Recently, we have witnessed the success of deep reinforcement learning (DRL) in many security applications, ranging from malware mutation to selfish blockchain mining. Like all other machine learning methods, the lack of explainability has been limiting its broad adoption as users have difficulty establishing trust in DRL models’ decisions. Over the past years, different methods have been proposed to explain DRL models but unfortunately, they are often not suitable for security applications, in which explanation fidelity, efficiency, and the capability of model debugging are largely lacking. In this work, we propose AIRS, a general framework to explain deep reinforcement learning-based security applications. Unlike previous works that pinpoint important features to the agent’s current action, our explanation is at the step level. It models the relationship between the final reward and the key steps that a DRL agent takes, and thus outputs the steps that are most critical towards the final reward the agent has gathered. Using four representative security-critical applications, we evaluate AIRS from the perspectives of explainability, fidelity, stability, and efficiency. We show that AIRS could outperform alternative explainable DRL methods. We also showcase AIRS’s utility, demonstrating that our explanation could facilitate the DRL model’s failure offset, help users establish trust in a model decision, and even assist the identification of inappropriate reward designs.