Agentic XAI Achieves 33% Better Explanations, Boosting Trust in AI Predictions

The challenge of translating complex artificial intelligence insights into easily understood explanations limits the widespread adoption of data-driven decision-making, particularly in fields requiring expert trust. Tomoaki Yamaguchi from Gifu University, Yutong Zhou from the Leibniz Centre for Agricultural Landscape Research, Masahiro Ryo from Brandenburg University of Technology Cottbus, and Keisuke Katsura from Kyoto University, address this problem by introducing a novel approach called Agentic Explainable Artificial Intelligence. Their research demonstrates how combining explainability techniques with large language models operating as autonomous agents significantly improves the quality and clarity of AI-generated explanations, moving beyond simple interpretation of data to actively refine understanding. By testing this framework with real-world rice yield data, the team reveals that iterative refinement enhances recommendation quality, but also highlights the importance of carefully balancing explanation depth with conciseness to avoid diminishing returns, establishing key principles for the design of effective agentic XAI systems.

AI, Agriculture, and Explainable Decision Support

Research increasingly focuses on applying artificial intelligence to agriculture, particularly using machine learning to predict crop yields, support precision farming, and assist smallholder farmers. A growing trend involves utilizing large language models not merely as tools, but as autonomous agents capable of independent action and interaction, exploring frameworks for building multi-agent systems. A crucial aspect of this work centers on explainable AI, aiming to make these complex models more transparent and trustworthy, especially in sensitive areas like agriculture and environmental science. The emphasis is on understanding why AI models arrive at specific predictions, with techniques developed to explain the outputs of any classifier.

Recent advances combine agentic AI with explainable AI, improving understanding and reasoning, while also addressing the ethical and societal implications of AI, including potential biases and sustainability concerns. These studies highlight the importance of responsible AI principles in agricultural applications, ensuring fairness and promoting sustainable practices. Specific areas of investigation include yield prediction using data from sources like unmanned aerial vehicles, development of decision support systems for farmers, and scaling AI-powered services to support smallholder operations. Researchers are also exploring the use of large language models to analyze extensive agricultural datasets, identifying new insights and patterns. The system employs an LLM as an autonomous agent, progressively improving explanations through repeated analysis and refinement cycles, applied to rice yield prediction. This involved constructing a framework with three interconnected components: XAI analysis of yield data, LLM-driven iterative refinement of recommendations, and systematic evaluation using both human experts and LLMs. The team utilized a dataset collected from 26 rice fields in Japan over three years, encompassing yield data, soil properties, rice varieties, cultivation practices, and meteorological data.

Yield maps were generated from unmanned aerial vehicle (UAV) imagery using a deep learning model, while soil properties were measured annually with a tractor-mounted sensor. The agentic XAI system performed SHAP analysis, identifying key yield drivers, and initiated an iterative refinement process spanning eleven rounds, generating enhanced farmer recommendations. This work combines SHAP-based explainability with a Large Language Model (LLM) operating as an autonomous agent, iteratively refining explanations through eleven rounds of analysis. The team tested this system using rice yield data gathered from 26 fields in Japan, focusing on complex interactions between soil, weather, and farming practices. Experiments revealed that the framework successfully improved recommendation quality, as assessed by both human crop scientists and LLM evaluators across seven key metrics.

The research demonstrates a critical bias-variance trade-off in XAI systems, showing that while early refinement rounds addressed a lack of explanation depth, excessive iteration led to diminished returns. Specifically, models lost 60-80% of their capabilities within a few cycles due to reward overoptimization and distribution shift, peaking at Rounds 3 and 4, indicating an optimal level of explanation comprehensiveness. This finding challenges the assumption of monotonic improvement in agentic XAI and provides evidence-based design principles for building trustworthy AI systems.

Agentic XAI Improves Rice Yield Predictions

This research demonstrates a new agentic explainable artificial intelligence (XAI) system that enhances the quality of agricultural recommendations, specifically for rice yield prediction. By combining SHAP-based explainability with large language models operating as autonomous agents, the team successfully generated explanations that improved upon initial recommendations, as assessed by both crop scientists and other large language models. Evaluations across multiple metrics revealed an average quality increase of 30 to 33 percent after just a few iterative refinement rounds. However, the study also identified a critical limitation, demonstrating that continuous refinement does not guarantee ongoing improvement. The team observed an inverted U-shaped trajectory, where explanations initially became more insightful but subsequently declined in practical utility due to increased verbosity and a lack of grounding in real-world context. This highlights the importance of strategic early stopping to prevent over-complication and maintain the usefulness of agentic XAI systems, with future work focusing on validating these principles in different agricultural contexts and integrating external knowledge sources.