Part 1: What Makes Good Post-Training Reasoning Data? From Theory to Efficient Data Curation
Speaker: Baharan Mirzasoleiman (Associate Professor at UCLA)

Part 2: Coming soon
Speaker:

November 18th
1-2pm CT
Location: Zoom & In Person at NOIRLab Tucson headquarters Main Conference Room (MCR)

Part 1 Abstract: Post-training has emerged as a critical stage for unlocking the reasoning capabilities of Large Language Models (LLMs). In particular, supervised fine-tuning (SFT) and reinforcement learning (RL) significantly shape how models perform multi-step reasoning, follow instructions, and generalize to new tasks. But, post-training performance is highly sensitive to the quality: diversity, and difficulty of the training data. First, I will discuss some recent theoretical work that provides insights into the characteristics of high-quality post-training data. Building on these insights, I will present principled and efficient algorithms for curating SFT and RL datasets that directly target examples which most effectively shape reasoning behavior, while dramatically reducing the cost of data creation.

Part 2 Abstract:

Speaker Bios:
Baharan is an Assistant Professor in the Computer Science Department at UCLA. Her research aims to make machine learning more sustainable, reliable, and efficient through principled improvements to how models are trained. She develops data-centric and optimization-guided methods that improve the quality of training data and the learning algorithms used throughout the foundation model lifecycle. Before joining UCLA, she was a postdoctoral research fellow at Stanford University and received her Ph.D. in Computer Science from ETH Zurich. She received an ETH medal for Outstanding Doctoral Thesis, was selected as a Rising Star in EECS by MIT, and received an NSF Career Award, a UCLA Hellman Fellows Award, an Okawa Research Award and an Apple Research Award.