Loop Engineering: constraining LLM hallucinations in environmental health reasoning with knowledge graphs

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models (LLMs) exhibit a marked tendency to hallucinate when inferring environmental exposure– biological response relationships. Although knowledge graphs partially address this issue, symbolic knowledge graph completion enables the derivation of environmental factors and associated rules, yielding candidate evidence for exposure–response associations. Nevertheless, a considerable fraction of the generated paths remain either extraneous to the domain or mechanistically ambiguous, thereby compromising the reliability and interpretability of LLM-augmented reasoning systems. Here we propose a knowledge graph augmentation framework grounded in Loop Engineering, which embeds an automatic filtering strategy designed to iteratively retain rules and evidence chains of explicit environmental and biological relevance, effectively constraining the LLM's generative space. We empirically validate the framework using a knowledge graph that harmonises wild bird occurrence data with ozone exposure records. Experimental results demonstrate that the framework, when combined with automatic path filtering, achieves a 60% reduction in hallucination rate and markedly enhances inferential interpretability. In comparison with conventional single-round knowledge graph retrieval, Loop Engineering imposes a more stringent constraint on LLM generation through iterative path quality optimisation. This framework offers a generalisable pathway toward trustworthy AI reasoning in environmental health and holds potential for extension to other LLM applications requiring rigorous knowledge-grounded inference.