AGEL-Comp: Neuro-Symbolic Framework for Compositional Generalization in AI Agents
The novel neuro-symbolic AI framework, AGEL-Comp, tackles the issues of compositional generalization in agents powered by LLMs. It combines a dynamic Causal Program Graph (CPG) serving as a world model, an Inductive Logic Programming (ILP) engine that generates Horn clauses based on experiences, and a hybrid reasoning core where sub-goals suggested by an LLM are validated through a Neural Theorem Prover (NTP). This setup facilitates a learning cycle of deduction and abduction, enhancing performance in interactive settings. Further details can be found in arXiv:2604.26522.
Key facts
- AGEL-Comp is a neuro-symbolic AI agent architecture.
- It addresses compositional generalization in LLM-based agents.
- The framework includes a dynamic Causal Program Graph (CPG).
- An Inductive Logic Programming (ILP) engine synthesizes new Horn clauses.
- A hybrid reasoning core uses an LLM and a Neural Theorem Prover (NTP).
- The system operationalizes a deduction-abduction learning cycle.
- The research is published on arXiv with ID 2604.26522.
- The work focuses on grounding actions in interactive environments.
Entities
Institutions
- arXiv