ToE Framework Enhances Fact-Checking with Explainable Evidence Trees

The proliferation of fake news and AI-generated misinformation, especially through "Generative Engine Optimization" (GEO) poisoning, poses a significant threat to information integrity and the reliability of LLM reasoning. To counter this, researchers have developed the Tree of Evidence (ToE), a novel framework for automated fact-checking that is both hierarchical and explainable. ToE conceptualizes each claim as a dynamically growing argument tree, allowing for a structured and iterative verification process. The framework integrates several key components: a reinforcement learning-driven agent responsible for retrieving evidence from multiple sources, an evidence evaluation agent that assesses the credibility and relevance of the retrieved information, and an argument tree aggregation algorithm that synthesizes the evidence to reach a conclusion. This iterative process allows claims to be decomposed, evidence retrieved, and then verified through a transparent and explainable chain of reasoning. A theoretical analysis of the retrieval process provides a formal error bound, guaranteeing that the learned policy converges close to the information-theoretically optimal policy. Experiments across various datasets and backbone LLMs demonstrated that ToE significantly outperforms competitive baselines, showing improvements ranging from 4 to 24 percentage points. These gains were particularly pronounced when dealing with adversarially poisoned inputs, highlighting ToE's robustness against sophisticated misinformation tactics.