DeepRAP: Deep Reasoning, Abstraction & Planning towards trustworthy Cognitive AI Systems

Expected Impact

The resulting portfolio will not only advance the scientific state-of-the-art but also build a robust, interoperable, and application-driven community, positioning Europe at the forefront of trustworthy cognitive AI. It should also lay the foundations for future European leadership in safe, human-centric cognitive AI, supporting sovereignty and competitiveness in key sectors. It will support the ambitions of the AI Act[1] and the European approach to Artificial Intelligence[2].

Expected Outcome

Ambitious proposals put forward under this call will deliver:

• Models and/or architectures that handle multimodal data and knowledge, uncertainty, and can be trained and deployed with constrained computational resources
• Provable trustworthiness mechanisms ensuring explainability, transparency, fairness, risk evaluation, security and alignment with ethical and legal standards, including fundamental rights and the EU AI Act, and
• Demonstrate the developed capabilities integrated in a cognitive AI system (reaching TRL4) performing complex real-world tasks (e.g., scientific discovery, decision support, problem solving) as well as simulations at a scale.

In addition, proposals will:

• Propose new methods and metrics for evaluating and certifying reasoning and trustworthiness in AI as well as the use of the computational resources
• Follow the FAIR principles ensuring all data, models, and results are Findable, Accessible, Interoperable, and Reusable to maximise transparency, reproducibility, and impact, and
• Develop synergies with EU initiatives such as TEFs (AI Testing and Experimentation Facilities)[3], eBrains[4], Resource for AI Science in Europe (RAISE)[5], AI-on-demand Platform (AIoD)[6] and the Quantum Flagship[7].

Portfolio approach

The composition of the portfolio of projects to be funded under the DeepRAP Challenge will ensure comprehensive coverage across the following categories with a view to ensuring breadth and enabling synergies between the projects:

• Category 1 – Cognitive Function Capability: Reasoning, abstraction, and planning should be covered by the selected portfolio.
• Category 2 – Technological Approach: The selected projects are expected to use a variety of technological approaches, including but not limited to, neuro-symbolic AI, deep learning, reinforcement learning, and novel frameworks inspired by interdisciplinary fields, and
• Category 3 – Use Case and Application Domain: The selected projects will cover a variety of real-world domains, such as industry, mobility, civil security, scientific discovery, health, cybersecurity, justice and human-robot interaction.

The selected projects will also be assigned to lead and/or engage in portfolio activities centred on the following priorities:

• Interoperability: Establishing common standards and protocols to ensure seamless alignment between projects
• Benchmark Development: Co-creating a DeepRAP benchmark with shared tasks and an open evaluation platform for transparent assessment
• Common Pilots: Delivering joint pilot demonstrations addressing complex real-world problems to showcase DeepRAP capabilities
• Multiagent Integration: where feasible, combining project outcomes into modular, multiagent AI systems demonstrating collective reasoning and planning through structured interactions among multiple agents
• Application Shaping: Defining impactful use cases and engaging stakeholders to guide the development and adoption of innovative cognitive AI systems, and
• Ethical and Societal Alignment: Proactively addressing ethical, legal, and societal considerations, including fundamental rights, transparency, privacy, safety, and fairness of cognitive AI systems.

Objective

Innovative ideas put forward under this Challenge must explore novel approaches, including combinations of existing techniques (i.e. neuro-symbolic AI), or the creation of entirely new frameworks that go beyond current, traditional, deep learning and reinforcement learning paradigms. These could be inspired by developments in diverse fields such as neuroscience, biology, physics, philosophy and more.

The proposals should address one or more of the following cognitive capabilities:

• Deep Reasoning: Moving beyond statistical pattern matching to support causal inference, logical reasoning, and context-aware or commonsense decision-making in complex, unstructured environments. This requires shifting from purely data-driven correlations to AI systems capable of understanding why patterns emerge, identifying underlying causes, and drawing valid conclusions through both deductive and inductive processes. Neuro-symbolic approaches, which combine the learning power of neural networks with the structured inference of symbolic reasoning are particularly encouraged to advance these capabilities. Integrating contextual and commonsense knowledge enables AI to interpret information more holistically, adapt decisions dynamically, and handle ambiguity and uncertainty. Deep reasoning systems should be able to reconcile multiple sources of information, provide transparent and explainable rationales for their outputs, and align with human values and expectations, ensuring trustworthy and accountable operations in demanding real-world scenarios.
• Deep Abstraction: Enabling AI systems to generalise insights from limited data by forming, manipulating, and refining high-level concepts, analogies, and representations that can be transferred across diverse application domains. This includes the development of internal world models to support abstraction, foster commonsense understanding, and integrate semantic and contextual awareness. Approaches that combine symbolic reasoning, analogical mapping, and representation learning are particularly encouraged, as they empower AI to interpret meaning, intent, and relationships within complex environments. Progress in deep abstraction is essential for achieving cognitive flexibility, robust transfer learning, and adaptive reasoning in dynamic, data-scarce, or rapidly evolving settings.
• Deep Planning: Developing robust, adaptive, and scalable planning algorithms/models capable of operating in open-world, agentic, or uncertain real-time environments. This involves leveraging advanced deep learning techniques such as deep reinforcement learning and architectures tailored for planning tasks to enable AI systems to autonomously devise, optimise, and adjust complex strategies in dynamic settings. Neuro-symbolic approaches integrating neural networks with symbolic reasoning are particularly encouraged to address uncertainty, provide formal guarantees, and enable explainable, dependable decision-making. Emphasis is placed on long-term, flexible planning that incorporates cognitive timing and predictive modelling, enabling systems to anticipate and adapt within dynamic contexts. Approaches should explore hierarchical planning across multiple temporal levels, contingency planning for effective fallback strategies, and continual re-planning to dynamically update plans as environments evolve. These advancements will underpin resilient, coordinated, and trustworthy AI planning in complex, unpredictable scenarios.

Scope

Artificial Intelligence (AI) systems have achieved remarkable progress as evidenced by the ability of Generative AI to recognise patterns and generate contextually relevant outputs based on ever larger models and associated datasets. However, despite the remarkable strides made over the past decade, there remains a significant gap between the capabilities of the human brain and machine intelligence, which must be overcome to achieve robust performance and enable effective interactions with users and stakeholders.

Current Generative AI models can release very accurate outputs and even solve some mathematical problems but might struggle with some complex reasoning benchmarks and to understand the real world. These models frequently fail to reliably solve logic tasks and long-term planning, even when provably correct solutions exist, limiting their effectiveness in critical applications where precision is essential.

Inspired by the human brain’s ability to process information at multiple levels of abstraction—enabling perception, reasoning, and goal-directed planning—the goal of this Challenge is to move beyond the current state-of-the-art in traditional AI approaches, whether symbolic (e.g., rules, decision trees, symbolic regression, etc.) or connectionist, neural (e.g., deep learning, large language models, reinforcement learning). The goal is to significantly improve the Reasoning, Abstraction, and Planning (RAP) capabilities of AI systems.

This will overcome the limitations of current deep learning models, which despite their strengths, have limitations in critical cognitive functions for abstraction, contextualisation, causality, explainability, and intelligible reasoning — competencies that are fundamental to move towards human-like intelligence.

[1] https://digital-strategy.ec.europa.eu/en/news/commission-launches-ai-innovation-package-support-artificial-intelligence-startups-and-smes

[2] https://ec.europa.eu/commission/presscorner/detail/en/ip_24_383 150 AI Act | Shaping Europe’s digital future (europa.eu)

[3] Sectorial AI Testing and Experimentation Facilities under the Digital Europe Programme | Shaping Europe’s digital future

[4] EBRAINS: Europe's Research Infrastructure for Brain Research - EBRAINS

[5]

[6] Home Page | AI-on-Demand

[7] Introduction to the Quantum Flagship | Quantum Flagship