Abstract: 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, Cong Zeng, Tatsuki Kuribayashi, Rishabh Bhardwaj, Bingchen Zhao, Yawen Duan, Yi Liu, Emad A. Alghamdi, Yaodong Yang, Yinpeng Dong, Soujanya Poria, Pengfei Liu, Zhengzhong Liu, Hector Xuguang Ren, Eduard Hovy, Iryna Gurevych, Preslav Nakov, Monojit Choudhury, Timothy Baldwin. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (System Demonstrations).
2025.
Introduction
In the current Large Language Model (LLM) landscape, development is often characterized as an "arms race" focused primarily on capabilities—reasoning, coding, and mathematical proficiency. Popular benchmarks like the Open LLM Leaderboard or Chatbot Arena reward these traits, often overlooking the critical dimension of safety. This has led to a significant gap in responsible AI development, where models may achieve top-tier performance while remaining vulnerable to biases, misinformation, and harmful prompt injections.
Libra-Leaderboard addresses this by shifting the focus from pure performance to a balance between capability and safety. By treating safety as a first-class citizen alongside utility, the framework incentivizes developers to build models that are both high-performing and ethically aligned.
Methodology and the Balance-Encouraging Metric
The Libra-Leaderboard framework consists of three main components: Libra-Eval, an automated backend for safety evaluations; the Safety Arena, an interactive platform for adversarial stress-testing; and a novel Balance-Encouraging Metric.
Traditional evaluation methods often use simple averages to combine safety and performance metrics. This is problematic because a model could be exceptionally intelligent but dangerously unsafe, yet still maintain a high average score. To solve this, the authors introduce a "distance-to-optimal" scoring method:
$$Score = 1 - \sqrt{\frac{(1-x)^2 + (1-y)^2}{2}}$$
In this formula, $x$ and $y$ represent normalized safety and capability scores. This approach rewards models that move toward the "ideal point" (1,1) and penalizes those that neglect one dimension to boost the other.
Evaluation Framework
The framework curates 57 safety datasets, with over 40 introduced after 2023 to minimize data contamination. These datasets cover four key dimensions:
- Direct Risky Prompts: Explicitly unsafe queries.
- Adversarial Attacks: Modified or misleading inputs designed to bypass guardrails.
- Instruction Hierarchy Attacks: Complex instructions (e.g., "ignore all previous instructions") that test model robustness.
- Over-sensitivity: Identifying when a model "over-refuses" harmless prompts, ensuring safety doesn't degrade basic utility.
To combat data leakage, the authors implement a quarterly update strategy, sampling new subsets for testing and releasing older instances to the public.
Key Results and Insights
The initial release evaluated 26 mainstream LLMs from 14 organizations. Models like Claude 3.5 Haiku, GPT-4o, and GPT-4-1106-preview emerged as leaders in the balanced ranking, successfully navigating the trade-off between power and safety.
Several key insights emerged from the evaluation:
- Organizational Recipes: Models from the same family (e.g., GPT-4 or Claude) show higher safety correlations than models from different organizations, suggesting that proprietary alignment data and organizational "recipes" are more influential than model size alone.
- Sustainability Gaps: Models generally resist direct risky prompts well but struggle with sophisticated adversarial attacks and instruction hierarchy tasks.
- Commercial vs. Open-Source: While commercial models currently lead in safety, the evaluation identified notable weaknesses in all state-of-the-art models, particularly regarding prompt extraction robustness.
Conclusion and Future Work
Libra-Leaderboard provides a necessary corrective to capability-only rankings. Future iterations will integrate human-centric data from the Safety Arena and expand to handle multi-modal safety (video, audio). The framework remains a living project, continuously sourcing fresh adversarial prompts to stay ahead of the evolving landscape of LLM vulnerabilities.
Summary updated