Assessment methodologies for General Purpose AI capabilities and risks (RIA) (AI/Data/Robotics Partnership)

• New assessment and validation methodologies enabling comprehensive evaluation of GPAI models and multimodal systems, including capabilities and associated risks.
• Research outcomes adopted by GPAI providers, policymakers, public institutions and other stakeholders for systematic evaluation of GPAI capabilities and risk profiles.
• Enhanced support to the EU AI Office in enforcing AI Act rules for GPAI models and enabling self-evaluation tools for developers to ensure compliance with AI Act requirements.

• Proposals must address one or more of the following research areas: (1) innovative capability elicitation and forecasting for GPAI models; (2) assessment of capabilities with high economic impact or misuse potential (including CBRN and cybersecurity risks); (3) interpretability/explainability methods for emergent AI capabilities; (4) design and publication of benchmark tests uncovering latent capabilities.
• Teams must be interdisciplinary, combining technical AI expertise with SSH disciplines to maximize societal relevance and impact.
• Full adherence to Horizon Europe Open Science requirements: open access to research outputs or clearly defined alternative access arrangements for GPAI providers, policymakers and stakeholders.
• Proposals must include robust progress-measurement mechanisms (qualitative/quantitative KPIs, benchmarking, progress monitoring), engagement in international contests, and use-cases demonstrating added value.
• Results must be shared via the AI-on-demand platform (or other relevant digital resources) to strengthen the European AI, Data and Robotics ecosystem.
• Mandatory allocation of tasks for cohesion with the ADRA Partnership and the Gen AI 4 EU central Hub (CSA HORIZON-CL 4-2025-03-HUMAN-18).
• Proposals should build on or collaborate with existing projects and develop synergies with other Horizon Europe and Digital Europe Programme initiatives.