Exploiting Language Models as a Source of Knowledge for Cognitive Agents

Abstract: LLMs provide capabilities far beyond sentence completion, including question answering, summarization, and natural-language inference. While many of these capabilities have potential application to cognitive systems, our research is exploiting LLMs as a source of task knowledge for cognitive agents, that is, agents realized via a cognitive architecture. We identify challenges and opportunities for using LLMs as an external knowledge source for cognitive systems and possible ways to improve the effectiveness of knowledge extraction by integrating extraction with cognitive architecture capabilities, highlighting with examples from our recent work in this area.

• The paper introduces a direct extraction approach that enables cognitive agents to query LLMs for targeted knowledge acquisition.
• The paper employs few-shot prompting and template-based queries to elicit context-aware responses from LLMs that are rigorously validated.
• The paper shows that integrating LLMs with cognitive architectures improves structured knowledge integration while reducing dependency on human oversight.

Introduction to Cognitive Agent Knowledge Acquisition

Cognitive architectures serve as significant tools in the development of intelligent systems — cognitive agents imbued with capabilities such as decision-making, planning, and learning. A recurrent challenge impeding their scalability has been the acquisition and integration of new knowledge to execute increasingly complex tasks in diverse domains. The potential of LLMs as a substantial knowledge repository offers a promising avenue but comes with its own set of challenges, such as data reliability and relevancy.

Integrating LLMs with Cognitive Architectures

This integration hypothesizes a synergistic relationship between the structural advantages of cognitive architectures and the extensive knowledge base of LLMs. Cognitive architectures contribute structured, context-aware processing abilities, while LLMs bring in the capacity for broad knowledge retrieval. The direct extraction of task knowledge from LLMs to cognitive agents can potentially bypass the limitations inherent in each approach when working in isolation.

Direct Extraction Approach

One proposed solution centers on direct extraction, where cognitive agents interact with LLMs to fulfill specific knowledge gaps. This entails the agent formulating queries tailored to context and processing the received LLM responses. Such interaction demands that responses from LLMs satisfy criteria such as interpretability, grounding to situational context, compatibility with agent affordances, and alignment with human expectations.

Implementation and Evaluation

Researchers have applied a step-wise method to leverage direct extraction, using a combination of template-based and few-shot prompting to elicit meaningful responses from LLMs. Verification of the knowledge obtained is crucial, involving processes like query refinement and human evaluation to ensure the applicability and correctness of the responses. The strategy utilized sees the agent identify knowledge gaps, prompt the LLM with tailored queries, analyze responses, and ultimately integrate verified knowledge.

Conclusion

Initial experimentation underscores the need for rigorous response evaluation to ensure reliable learning. The analysis also suggests that integrating various knowledge sources within task learning could minimize dependencies on human oversight, indicating a fruitful pathway for cognition-driven systems. The full realization of this integrated approach, encompassing all facets of task learning, remains a rich field for future exploration.