Reliable Problem Solving via LLMs with Symbolic Argumentation
Aus International Center for Computational Logic
Reliable Problem Solving via LLMs with Symbolic Argumentation
Vortrag von Antonis Kakas
- Veranstaltungsort: APB-2026
- Beginn: 30. April 2026 um 11:00
- Ende: 30. April 2026 um 12:00
- Forschungsgruppe: Computational Logic
- Event series: Research Seminar Logic and AI
- iCal
In modern AI the aim of a system is to address the problem requirements directly from the natural language description of the problem. Clearly, such a problem specification is expressed under human reasoning and hence a problem solving system would be advantageous if it could also operate under a formalism that is close to human reasoning. The recent development of Large Language Models (LLMs) offers such a possibility as these systems appear to handle natural language in a human-like fashion. But their reliability is under question as is their ability to reason underneath their process of constructing well-formed sentences in order to explain and argue for their solutions to problems. The question the arises: do we develop and train LLMs as reasoners or do we connect these to other formal (symbolic) reasoners who do not suffer from the above shortcomings? In this talk, the speaker will present a neuro-symbolic framework, based on the integration of the Natural Language capabilities of LLMs with Argumentation-based Reasoning. In this approach, instead of the LLM carrying out the reasoning, this is delegated to COGNICA, a Cognitive Argumentation system, which carries out the reasoning within a Controlled Natural Language. Specifically, LLM-COGNICA is a hybrid neural-symbolic framework for reasoning under a decision policy specified within Natural Language and executed via Explainable Argumentation. This offers the reliability of formal problem solving and allows the human developer and problem solver to be in control of the system development process. The talk will present the Argumentation foundations and Argumentation Technology underlying the neuro-symbolic integration of the LLM-COGNICA framework and describe how this can be employed to build reliable real-life application systems.
