Exciting to have our team playing a role at NeurIPS this year, spanning LLM alignment, graph foundational models, agent evals, healthcare AI and more. We are incredibly proud of Steven (Jiayuan) Ding, PhD for his leadership. That’s a wrap, 2026.
NeurIPS 2026 has wrapped up. Grateful to share that five pieces of work I had the privilege of contributing to made it in this year. Research Papers: 1. Crafting Reversible SFT Behaviors in Large Language Models Supervised fine-tuning teaches LLMs new behaviors, but these behaviors are often difficult to isolate or control once they are baked into the model. We study how to make SFT-induced behaviors reversible at inference time, enabling more flexible control over model behavior without retraining or changing model weights. Paper link: https://lnkd.in/gNhB9FSd 2. Understanding Graph Self-Supervised Pre-training under Distribution Shifts: A Scaling Law Perspective Graph self-supervised pre-training has become a standard recipe, but it remains unclear when pre-training actually transfers under distribution shifts. We analyze this problem through a scaling-law perspective, helping characterize when graph pre-training improves generalization and when it breaks down. Paper link: https://lnkd.in/g3RbwDva 3. I’m also glad to contribute to Agents’ Last Exam, a large-scale community benchmark developed with 250+ industry experts to evaluate AI agents on long-horizon, economically valuable, real-world professional tasks with verifiable outcomes. The benchmark spans 1,000+ tasks across 55 subfields and 13 industry clusters, aiming to measure whether agents can move beyond benchmark success toward real-world productivity. Paper link: https://lnkd.in/gqf_JxDg 4/5: I’m also excited to have led two broader NeurIPS efforts this year: The Virtual Embryo Challenge, a competition on generative modeling of embryogenesis across space, scale, and time, and The Third Workshop on GenAI for Health, focused on agentic systems, clinical trust, and the future of healthcare AI. Competition: https://lnkd.in/gVCYQP4k Workshop: https://lnkd.in/gDpP53tH None of this happens alone. Huge thanks to every co-author, especially the students who carried so much of this work. This year spanned LLM alignment, graph foundation models, agent evaluation, computational biology, and healthcare AI, and it was a powerful reminder of why I love working at the intersections. Yuping Lin, Pengfei He, Yue Xing, Yingqian CUI, Hui Liu, Zhen Xiang, Bingheng Li, Yiyou Sun, David (Xinyang) Han, Weize Xu, Siyu He, James Zou, You He, Eric Xing, Le Song, Pranav Rajpurkar, Junyuan Hong, Ehsan Adeli, and so many others who made this year possible. And a special thank-you to the people who've shaped how I think and work: Munjal Shah and Subhabrata (Subho) Mukherjee for the trust and room to pursue this at Hippocratic AI, and Ying Ding, Jiliang Tang, and Xiaojie Qiu for years of mentorship and collaboration.