Bowen Wei
Hello! My name is Bowen Wei, and I am a fourth-year Ph.D. student in Computer Science at George Mason University. I am fortunate to be advised by Professor Ziwei Zhu.
My research sits at the intersection of trustworthy AI and LLM reasoning, with a growing focus on agentic AI safety and reinforcement learning for building reliable, autonomous systems. I develop interpretable learning methods—including prototype-based and neuro-symbolic approaches—that make model decisions transparent, auditable, and robust for high-stakes applications, revealing what evidence or rules drive a prediction and when the model may fail.
In parallel, I investigate agent safety evaluation and reinforcement learning techniques for agentic reasoning, including safety benchmarking under adversarial prompt injection and self-play frameworks. Ultimately, I aim to build AI systems that are both interpretable in their inner logic and safe and reliable in autonomous, evidence-grounded decision-making.
News
| Nov 03, 2026 | 🎉 Our paper “Interpretable Classification via a Rule Network with Selective Logical Operators” has been accepted to the Main Conference at WSDM 2027! |
|---|---|
| Oct 20, 2026 | 🎉 Our paper “A Logical-Rule Autoencoder for Interpretable Recommendations” has been accepted to IEEE BigData 2026! |
| Aug 24, 2026 | 🎉 Our paper “COSE: Confidence-Orchestrated Self-Evolution for Effective LLM Reasoning” has been accepted to the Main Conference at EMNLP 2026! |
| Apr 14, 2026 | 🎉 Two papers accepted to ACL 2026! “VIGNETTE: Socially Grounded Bias Evaluation for Vision-Language Models” as an oral in the Main Conference, and “Context-Aware Decoding for Faithful Vision-Language Generation” in Findings. |
| Nov 08, 2025 | 🎉 Our paper “Making Sense of LLM Decisions: A Prototype-based Framework for Explainable Classification” has been accepted for an oral presentation at AAAI 2026! |
Selected Publications
- ICLR 2027Under ReviewHILT: Hazard-Informed Localized Teaching for Safer Language Models2026
- arXivPreprint
- WSDM 2027MainInterpretable Classification via a Rule Network with Selective Logical OperatorsIn Proceedings of the 20th ACM International Conference on Web Search and Data Mining, 2027Main Conference acceptance rate: 15.8% (277 / 1,756).
- EMNLP 2026MainCOSE: Confidence-Orchestrated Self-Evolution for Effective LLM ReasoningIn Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, 2026Main Conference acceptance rate: 15.4% (2,719 / 17,669).
- IEEE BigData 2026A Logical-Rule Autoencoder for Interpretable RecommendationsIn Proceedings of the 2026 IEEE International Conference on Big Data, 2026* Equal contribution. Acceptance rate: 19.2% (201 / 1,046).
- ACL 2025Main
- NeurIPS LAW 2025
- ACL 2026OralVIGNETTE: Socially Grounded Bias Evaluation for Vision-Language ModelsIn Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, Jul 2026Selected as an SAC Highlight at ACL 2026.
- ACL 2026FindingsContext-Aware Decoding for Faithful Vision-Language GenerationIn Findings of the Association for Computational Linguistics: ACL 2026, Jul 2026
- ICML 2026MainKnowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron EnhancementJul 2026
- WACV 2026MainCAAC: Confidence-Aware Attention Calibration to Reduce Hallucinations in Large Vision-Language ModelsJul 2025