Data Science Talk Series 

Advancing Instruction Following and Steerability for LLMs

Abstract: This talk explores the critical role of Instruction Following (IF) and steerability as core capabilities of Large Language Models (LLMs). It addresses contemporary challenges including multi-turn dialogue constraints, system prompt adherence, and the evaluation of open-ended or qualitative instructions. The session introduces two benchmarking frameworks: Multi-IF (designed for multi-turn and multilingual rule-based instructions) and AdvancedIF (centered on expert-curated, rubric-based evaluation). Furthermore, it presents the RIFL (Rubric-based Reinforcement Learning) framework, a post-training pipeline that leverages rubric generation and reward shaping to enhance model performance while mitigating the common issue of "reward hacking."

Bio of the speaker: Yun He is a Senior Research Scientist at Meta Superintelligence Labs (MSL), where he leads research initiatives on LLM post-training. He holds a Ph.D. from Texas A&M University, where his research concentrated on NLP and recommender systems. His current research interests lie at the instruction following, reward modeling and RLHF.