Events

Global AI Frontier Lab: Seminar Series: Scaling Test-Time Compute for LLMs: Adaptation, Cognitive Behaviors, and Verification

Speaker: Jack Lu

Location: 1 MetroTech Center, Room floor 22

Date: Monday, March 23, 2026

We are pleased to announce the next session of the Global AI Frontier Lab: Seminar Series on March 23rd, 2026. Jack Lu will be presenting “Scaling Test-Time Compute for LLMs: Adaptation, Cognitive Behaviors, and Verification”. Dinner & networking will begin at 6:00 PM and the seminar will start at 7:00 PM EST. The seminar will be held at the Global AI Frontier Lab at 1 Metrotech Center, Brooklyn, NY 11201. This event will be in-person & online. In-person attendance is strongly encouraged for Lab researchers in NYC. All attendees must RSVP to participate. For online attendees, a Zoom link will be sent out prior to the event. Please reach out to global-ai-frontier-lab@nyu.edu with any questions. We hope to see you there!

Abstract: Scaling test-time compute is emerging as a powerful axis for improving LLM performance beyond pretraining and fine-tuning. This talk presents three works that each explore this from a different angle. First, Context Tuning spends test-time compute on adapting how in-context examples are represented, achieving strong few-shot performance without updating model weights. Second, SkillFactory teaches LLMs to self-improve by distilling cognitive behaviors like verification, retry, and reflection from their own outputs, priming them to productively use extra tokens during and after post-training. Finally, we take a systematic look at LLM verification: when does spending test-time compute on having a model judge its own or another model's outputs actually lead to improvement? We introduce principled metrics and reveal how model family, post-training, and task type determine whether and when verification yields meaningful gains.

Bio: Jack Lu is a third-year Computer Science PhD student at New York University, advised by Mengye Ren and supported by the NSERC CGS-D Scholarship. He is interested in improving and understanding the reasoning and adaptation capabilities of large language models, with a focus on test-time scaling and self-improvement. This summer, he will join NVIDIA as a research intern working on reasoning VLA and world models. Previously, Jack interned at Waabi and NVIDIA with Raquel Urtasun and Sanja Fidler, working on generative models for synthetic data generation. He received his BMath in Computer Science and Mathematics from the University of Waterloo.