Hybrid Fact-Checking that Integrates Knowledge Graphs, Large Language Models, and Search-Based Retrieval Agents Improves Interpretable Claim Verification

A modular fact-checking pipeline that combines knowledge graphs, large language models, and web search agents has demonstrated strong performance on a standard benchmark without task-specific fine-tuning, according to new research [1]. The system, described in a paper by Timo Cavelius, achieves an F1 score of 0.93 on the FEVER benchmark’s Supported/Refuted split [1][2]. It operates through three autonomous steps: a Knowledge Graph Retrieval stage that performs rapid one-hop lookups in DBpedia, an LM-based classification guided by a task-specific labeling prompt, and a Web Search Agent that activates only when knowledge graph coverage is insufficient [1][2]. The pipeline is open-source and designed with fallback strategies intended to support generalization across datasets [1][2]. Large language models can produce fluent text but often lack grounding in verified information, while knowledge-graph-based fact-checkers offer precise and interpretable evidence at the cost of limited coverage or higher latency [1][2]. The hybrid approach attempts to leverage the strengths of each component, using the knowledge graph for quick, structured retrieval and the language model for classification with internal rule-based logic [1][2]. A targeted reannotation study examined claims originally labeled as Not Enough Information (NEI) in the FEVER dataset. The authors report that their approach frequently uncovers valid evidence for these claims, a finding confirmed by both expert annotators and LLM reviewers [1][2]. The paper was first submitted on 5 November 2025 as a 195 KB preprint and updated on 26 June 2026 with a 9,519 KB version [1]. Fact-checking research has increasingly focused on interpretability and evidence retrieval. The FEVER benchmark, which stands for Fact Extraction and VERification, provides a standardized testbed for evaluating claim verification systems. The new pipeline’s modular design allows individual components to be swapped or upgraded, and the reliance on DBpedia as a structured knowledge source offers a transparent evidence trail that contrasts with purely generative approaches [1][2].