SEAL: Self-Evolving Agentic Learning for Conversational Question Answering over Knowledge Graphs

SEAL introduces a two-stage semantic parsing framework that combines large language models with agentic learning to improve conversational question answering over knowledge graphs. The system self-evolves through dialog history and execution feedback without retraining, achieving state-of-the-art results on complex multi-hop reasoning and aggregation tasks while reducing computational costs.

SEAL represents a meaningful advancement in how AI systems process complex questions against structured knowledge bases. The framework addresses a fundamental challenge in semantic parsing: converting natural language queries into executable logical forms that accurately represent user intent while maintaining syntactic validity. This matters because knowledge graph question answering powers enterprise search, virtual assistants, and data retrieval systems where accuracy directly impacts user trust and operational efficiency.

The innovation centers on a hybrid approach combining LLM extraction with agentic refinement rather than relying on end-to-end neural models that frequently generate semantically misaligned queries. By decomposing the problem into a minimal semantic core followed by template-guided completion, SEAL reduces hallucination risks while maintaining computational efficiency. The self-evolving mechanism—integrating memory modules and reflection—enables continuous improvement from real dialog interactions, mimicking how humans learn from conversational context.

For practitioners building conversational AI systems, SEAL's approach offers practical improvements in handling multi-hop reasoning (questions requiring multiple logical steps) and aggregation queries (requiring mathematical operations across results). The elimination of explicit retraining cycles accelerates adaptation to domain-specific knowledge graphs, reducing deployment friction for enterprise applications. The framework's demonstrated performance gains on the SPICE benchmark suggest broader applicability across different knowledge domains.

Looking ahead, the scalability of this approach to real-world enterprise knowledge graphs remains a critical validation point. Integration with retrieval-augmented generation systems and exploration of how self-evolution performs on out-of-distribution queries will determine whether SEAL's promise translates to production environments.

• →SEAL combines LLM-based extraction with agentic calibration to reduce syntactically invalid logical forms in knowledge graph queries.
• →Self-evolving mechanism learns from dialog history and execution feedback without requiring model retraining.
• →Achieves state-of-the-art performance on multi-hop reasoning, comparison, and aggregation tasks.
• →Two-stage semantic parsing approach improves both structural accuracy and computational efficiency.
• →Framework addresses scalability challenges in entity-relation linking over large knowledge graphs.