Hi, I'm
Tiankai (Raymond) Yang
I’m a third-year PhD student in Computer Science at the University of Southern California, advised by Prof. Yue Zhao in the FORTIS Lab. I earned an MS in Machine Learning & Data Science from USC and a BE in Software Engineering from Nankai University.
My research focuses on the Trustworthiness of Large Language Models and Agents. Specifically:
- Post-training Alignment for Trustworthy LLMs — Safety Alignment, Preference Learning
- Robust and Reliable LLM Inference — Hallucination mitigation, Jailbreak / OOD Detection, Multimodal Robustness, Model Selection & Routing
- Trustworthy LLM Agents and Agentic Systems — Agent Safety, Runtime Reliability, Multi-Agent Orchestration
In summer 2026, I interned at LinkedIn as an AI/ML Engineer Intern (GenAI), working on user simulation agents, LLM reasoning, and post-training with GRPO and OPSD.
News
- Aug 2026 🎉 AcceptedDOG-DPO: Dynamic Optimization in Geometry for Safety Alignment accepted to EMNLP 2026 Findings. arXiv.
- Aug 2026 📝 PreprintHyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory is now on arXiv.
- Jul 2026 🎉 AcceptedFlexRouter: Learning Complementary Model Sets for Flexible LLM Routing accepted to COLM 2026.
- May 2026 🏆 AwardCoAct: Co-Active LLM Preference Learning with Human-AI Synergy selected for an Oral presentation at ACL 2026. ACL Anthology.
- Apr 2026 📝 PreprintCat-DPO: Category-Adaptive Safety Alignment is now on arXiv.
- Apr 2026 📝 PreprintNo Attacker Needed: Unintentional Cross-User Contamination in Shared-State LLM Agents is now on arXiv.
- Feb 2026 💼 InternshipI will join LinkedIn as an AI/ML Engineer Intern, Generative AI this summer. See you in Sunnyvale!
- Jan 2026 🎉 AcceptedCoAct: Co-Active LLM Preference Learning with Human-AI Synergy accepted to ACL 2026. ACL Anthology.
- Aug 2025 🎉 AcceptedAD-AGENT: A Multi-agent Framework for End-to-end Anomaly Detection accepted to IJCNLP-AACL 2025 Findings. ACL Anthology.
- May 2025 🎉 AcceptedAD-LLM: Benchmarking Large Language Models for Anomaly Detection accepted to ACL 2025 Findings. ACL Anthology.
- Feb 2025 🎉 AcceptedDPU: Dynamic Prototype Updating for Multimodal Out-of-Distribution Detection at CVPR 2025 as a Highlight. Poster.