[论文] VERA: A Responsible AI Dashboard for Assessing LLM Deployments in Ligh...

论文概要

研究领域: ML 作者: Robin Quériaux, Solveig Bachellery, Léa Déléris, Saurabhsingh Rajput, Mariam Barry, Tushar Sharma, Jérôme Lebecq 发布时间: 2026-10-22 arXiv: 2610.31588

中文摘要

在受监管环境部署LLM的机构需要负责任AI实践的技术证据。COMPL-AI提供EU AI Act的技术解读，将证据组织为六个原则组和十八项技术要求，其中十二项可由现有基准衡量。VERA是部署于一家大型欧洲银行的开源自托管仪表板：审查者选择模型并运行公共基准，VERA为十二项可衡量要求评分并记录每次评估的能源消耗。仪表板将评估时间从56秒降至平均30秒，组织内调查实用性评分4.36/5。仪表板和学习材料已公开发布以促进不同组织环境中的可复制性。

原文摘要

Institutions deploying LLMs in regulated settings need technical evidence of responsible AI practices before approving deployment. COMPL-AI provides a technical interpretation of the EU AI Act organizing evidence under six principle groups and eighteen technical requirements, twelve measurable by current benchmarks. We present VERA, an open-source self-hosted dashboard deployed at a large European bank. A reviewer selects a model and runs public benchmarks; VERA scores the twelve requirements and instruments energy consumption for environmental impact. The dashboard reduced assessment time from 56 seconds using raw files to 30 seconds on average, and scored 4.36/5 on usefulness.

#论文 #arXiv #ML #小凯