A Community-Driven Vision for a New Knowledge Resource for AI
This paper synthesizes the perspectives of over 50 researchers regarding the persistent deficit in general-purpose, verifiable, and open knowledge resources in AI, a gap left unresolved despite the foundational efforts of projects such as Cyc, WordNet, and commercial knowledge graphs. The authors argue that a new, openly-engineered knowledge infrastructure is vital, and propose both the technical and social scaffolding necessary to build and sustain it.
Motivation and Context
Despite the proliferation of LLMs and the commercial impact of entity-centric knowledge graphs (e.g., Google’s Knowledge Graph), AI still lacks comprehensive, trusted, and inspectable repositories of world and commonsense knowledge. The absence of such resources diminishes AI reliability in domains requiring explicit, contextual, and formally represented knowledge—particularly in robotic planning and fact verification.
The Cyc project’s original ambition of encoding everyday commonsense remains only partially realized. The authors highlight practical bottlenecks confronting the current generation of AI systems:
- LLMs encapsulate knowledge implicitly, which impedes transparency, formal guarantees, and error inspection.
- Knowledge graphs narrowly focus on named entities and lack the expressivity to model general rules, causality, or nuanced relationships.
- Custom, ad hoc knowledge bases are prevalent in applied fields (e.g., robotics, biomedicine), leading to duplication of effort and silos of domain competence that lack interoperability.
Requirements for a Modern Knowledge Resource
The paper stipulates several core requirements and distinct design shifts for a new resource:
- Expressiveness: Beyond RDF/OWL triples, the resource should accommodate richer representation languages—such as Answer Set Programming (ASP), Rulelog, and other logic programming dialects, as well as probabilistic and causal models.
- Openness and Community Contribution: The infrastructure must be open-source, with permissive licensing to enable widespread adoption and adaptation. Community curation and social conventions are essential for maintaining quality and relevance.
- Provenance and Contextualization: Knowledge statements should be paired with provenance information, contextual applicability, and, when possible, multimodal evidence.
- Interoperability: The resource should provide APIs and structures to facilitate integration into agentic workflows, LLM augmentation, robotics, and other emerging application paradigms.
Foundational and Domain-Specific Knowledge
The authors delineate the stratification of knowledge into foundational (e.g., time, space, causality, actions) and domain-specific levels (e.g., medical, legal, physical sciences). Key arguments emphasize:
- Reuse of Existing Ontologies: Prior work in qualitative spatial/temporal reasoning, formalized psychology, and constraint/causal models offers a springboard for bootstrapping the knowledge base.
- Systematic Curation Methodology: A scalable approach is needed, combining corpus analysis, iterative formalization, and a continuous loop of refinement and evaluation, potentially leveraging LLMs to elicit and formalize domain knowledge.
Open Problems in Knowledge Formalization
The paper identifies enduring challenges:
- Multi-modal, cross-domain interoperability (e.g., linking spatial, linguistic, and causal knowledge).
- Disambiguating and translating between heterogeneous representations.
- Intrinsic and extrinsic evaluation of knowledge modules for theoretical soundness and real-world task competence.
Automated Reasoning Integration
The resource must be paired with robust and scalable reasoning capabilities:
- Deductive, Inductive, and Abductive Reasoning: Support for diverse reasoning modes is necessary to accommodate the full range of AI tasks, from formal proof generation to human-like hypothesis formation and belief revision.
- Contextual Reasoning: Localized microtheories, context-aware inference, and viewpoint selection are required for robust decision-making in dynamic environments.
- Human Reasoning Modeling: While direct emulation of human fallibility is avoided, systems should augment and correct human-like reasoning, incorporating judgment calls, relevance filtering, and counterfactual capability.
Human-in-the-Loop and Automated Curation
Curation strategies blend human expertise with automation:
- Human Oversight Remains Central: For tasks such as schema design and source selection, expert input is indispensable.
- Role of LLMs: LLMs are positioned as assistants in knowledge acquisition, formalization, and user interaction—capable of surfacing implicit knowledge, aiding with translation to formal semantics, and lowering the barrier for community contribution.
- Continuous Feedback Loop: The optimal model leverages humans to specify, test, and refine knowledge structures, while automation handles laborious formalization and validation steps.
Educational and Cultural Initiatives
The authors assert that deficiencies in the education pipeline and prevailing attitudes toward knowledge representation must be addressed for long-term sustainability:
- Curricular Gaps: Courses on knowledge representation are limited, outdated, and insufficiently grounded in real-world applications. Integration throughout computer science and related disciplines is advocated.
- Modular, Plug-and-Play Teaching Resources: Teaching modules, sandboxes, and bodies of knowledge—akin to the software engineering SWEBOK—should be broadly deployed to lower adoption costs and standardize best practices.
- Early Logic Exposure: Introducing logic and reasoning at pre-university levels is highlighted as a way to nurture the next generation of knowledge engineers.
Evaluation Methodologies
The paper critiques current benchmark-driven AI evaluation, citing proxy failures and the inadequacy of task-based metrics. It recommends:
- Expert-interview-based Evaluation: Enlisting domain and AI experts for hands-on system assessments.
- Virtual and Game Environments: Using simulated, constrained environments to rigorously test reasoning abilities.
- Process Transparency: Emphasizing evaluations that examine not only results but the sequence of knowledge and reasoning steps, leveraging test sets that reflect real task diversity and edge cases.
Implementation and Community Governance
To operationalize this vision:
- A Dynamic, Open Repository: Inspired by the Hugging Face model, the resource should offer modular, task-oriented knowledge packages, sandboxes for experimentation, and seamless installation for practical use.
- Nonprofit Stewardship: A foundation structure with academic and commercial representation is proposed to ensure neutrality, inclusivity, and long-term viability.
- Emphasis on Practical Use Cases: Initial focus areas include LLM augmentation, agentic AI, robotics, and fact-checking, which both stimulate adoption and probe expressivity.
Implications and Future Directions
The authors argue for a renewed investment in curated, explicit knowledge as a counterbalance to the limitations of deep learning-centric AI. They predict that re-integrating explicit, inspectable knowledge will unlock new dimensions of robustness, interpretability, and compositional generalization across AI systems. The work foregrounds the critical interplay of engineering methodologies, open social processes, and sustained educational reform in catalyzing the next wave of knowledge-driven AI.
Notable Claims
- Survey data indicates a strong belief among AI researchers in the necessity of symbolic methods for human-level reasoning, with over 60% estimating that at least half of reasoning needs explicit knowledge.
- Current LLMs and knowledge graphs, while effective in conventional applications, are structurally incapable of providing formal guarantees, transparency, and systematic extensibility required by mission-critical or introspective AI systems.
Outlook
This synthesis outlines a roadmap and posits that real progress depends on the convergent evolution of technical, social, and educational ecosystems. The success of the envisioned infrastructure hinges not only on technical innovation but on the iterative, community-driven curation, open governance, and integration into practical, high-impact applications spanning the breadth of AI.