SafetyAccountabilityTransparencycs.CL
Libra-Leaderboard: Towards Responsible AI through a Balanced Leaderboard of Safety and Capability
Haonan Li, Xudong Han, Zenan Zhai, Honglin Mu, Hao Wang, Zhenxuan Zhang, Yilin Geng, Shom Lin, Renxi Wang, Artem Shelmanov, Xiangyu Qi, Yuxia Wang, Donghai Hong, Youliang Yuan, Meng Chen, Haoqin Tu, Fajri Koto, Tatsuki Kuribayashi, Cong Zeng, Rishabh Bhardwaj, Bingchen Zhao, Yawen Duan, Yi Liu, Emad A. Alghamdi, Yaodong Yang, Yinpeng Dong, Soujanya Poria, Pengfei Liu, Zhengzhong Liu, Xuguang Ren, Eduard Hovy, Iryna Gurevych, Preslav Nakov, Monojit Choudhury, Timothy Baldwin
December 24, 2024
arXiv.org
4 citations
Key Takeaways
-
Current AI models have significant safety gaps: The evaluation of 26 mainstream LLMs from 14 leading organizations revealed "critical safety challenges even in state-of-the-art models," indicating that top-performing AI systems are not automatically safe systems.
-
Traditional ranking methods create perverse incentives: By averaging performance and safety metrics, conventional leaderboards can reward models that excel in capability while neglecting safety—Libra-Leaderboard's "distance-to-optimal-score method" instead incentivizes balanced development across both dimensions.
-
Safety and capability should be jointly optimized, not traded off: The framework encourages developers to pursue "joint optimization of capability and safety" rather than treating these as competing goals, suggesting responsible AI requires simultaneous progress on both fronts.
-
Transparent, balanced evaluation drives accountability: The combination of a "dynamic leaderboard with an interactive LLM arena" creates visibility into how models perform on safety alongside capability, making it harder for organizations to hide safety shortcomings behind strong performance metrics.
-
How we measure AI matters for responsible development: The paper demonstrates that evaluation frameworks themselves shape developer incentives—using a balanced ranking system rather than traditional averaging can fundamentally change which models get built and deployed.