Description: The term foundation model describes complex neural networks that have been trained on large amounts of data and are characterized by a high degree of generality and versatility in their possible applications. For this purpose, the models have a large number of parameters (several hundred billion) that are learned during training. The implicit knowledge acquired from the data is represented in the parameters and thus represents the basis (foundation) for further possible applications. Generative foundation models differ from previous AI models in that they generate complex text, image or audio content (or a combination of these) on the basis of free input (prompts) and can therefore be used for a wide range of complex tasks. The initial training effort for such models is extremely high and therefore time- and resource-intensive.
With the paradigm shift from mostly unimodal models with narrowly defined functionality to generative models with self-monitored learning of huge multimodal data volumes, the field of application of the technology has been significantly expanded. At the same time, however, evaluation has also become much more complex. While measurements using quantitative approaches are possible for classic regression, classification and structure recognition tasks, qualitative approaches covering multiple property dimensions are necessary for complex generative tasks. Problems such as the invention of facts and sources in large language models (“hallucinating”), the possibility of manipulatively inducing unintended model behavior (through direct or indirect prompt injection), the perpetuation of stereotypes from the training data in the model output (bias) and the high computing power required for inference make a holistic evaluation all the more necessary if use in sub-domains of internal and external security is envisaged. Previous approaches are not sufficiently suitable for use in a German-speaking context and in consideration of security domain-specific tasks.
Against this background, the aim of the procedure is to provide innovative research and development services in a competitive framework (“challenge”) for the holistic evaluation of generative foundation models in a security context.
Awarding body: Agentur für Innovation in der Cybersicherheit GmbH
Services and products: Research and development services and related consulting
Tender scope: EU member states, Switzerland, NATO member state or NATO partner, in particular AP4 (Asia-Pacific Partner – South Korea, Japan, Australia and New Zealand)
Procurement procedure: Pre-commercial procurement
Type of award: Based on a negotiated procedure with a call for competition
Deadline for submission of requests to participate + outline: 31.07.2025, 10:00 a.m., via www.evergabe-online.de or directly https://ted.europa.eu/de/notice/-/detail/358520-2025
Place of fulfillment: Halle (Saale), district-free city
Contact: only via www.evergabe-online.de
Contract notice number TED: 358520-2025
CPV code: 73000000
More information on the research program: https://www.cyberagentur.de/programme/hegemon/