Conferencia | GFlowNets: Generación Diversa para el Razonamiento Matemático en LLM

We are pleased to invite you to an interactive online conference on February 25 at 12:30 p.m. at the UEMF Conference Center, Yellow Room. The conference will be delivered by Prof. Salem Lahlou, Professor of Machine Learning at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).

The presentation, entitled “GFlowNets: Diverse Generation for Mathematical Reasoning in LLMs”, will explore recent advances in generative modeling for structured and compositional objects. GFlowNets are designed to sample solutions proportionally to a reward function, enabling the generation of diverse, high-quality candidates rather than converging to a single optimal solution.

The talk will first present the mathematical foundations of GFlowNets, including flow-matching principles, training objectives, and continuous extensions. It will then focus on a detailed application for fine-tuning large language models in mathematical reasoning. By combining step-level GFlowNets with automated Process Reward Models, this approach achieves a 9.4% improvement in out-of-distribution generalization on SAT MATH, significantly outperforming Proximal Policy Optimization (PPO) while preserving solution diversity.

The presentation will conclude with a brief overview of how these methods can be extended to other domains, such as biological sequence design, highlighting the broader impact and versatility of the GFlowNet framework.