✨ That's a wrap on the 2025 #ACM FAccT conference! We're thrilled to have brought together researchers, practitioners, and innovators from around the globe to explore the latest advances in fairness, accountability, and transparency in sociotechnical systems.

A huge thank you to all our attendees, presenters, sponsors, and organizers for making #FAcct2025 a resounding success! 👏

Until next time, keep pushing the boundaries of equitable tech for all 🚀

Today is our last day at the #FAccT2025 conference 💔

As our time in Athens draws to a close, this year's program chairs would like to share some reflections on the FAccT 2025 review process in a new blog post.

https://facct-blog.github.io/2025-06-26/review-process

Reward models stand in for human values when aligning LLMs using reinforcement learning with human feedback (RLHF). What values do these models actually encode and how do they compare to independent measures of human values?

Led by my colleague Brian Christian, at #FAcct2025 we present a novel approach to reward model interpretability to answer these questions.

We found wild disagreement among different models and a slew of biases and interesting patterns which raise questions about how reward models are trained and used.

Reward Model Interpretability via Optimal and Pessimal Tokens
by
Brian Christian, Hannah Rose Kirk, Jessica A.F. Thompson, Christopher Summerfield, Tsvetomira Dumbalska

https://arxiv.org/abs/2506.07326