ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation

A new automated fact-checking framework called Tree of Evidence (ToE) models claims as dynamically expanding argument trees to counter AI-generated misinformation, according to a paper published on arXiv. The system integrates a reinforcement learning agent for multi-source retrieval with an evidence evaluation and aggregation algorithm. [1] The framework is designed to address the contamination of large language model (LLM) reasoning by adversarially crafted content surfaced through Generative Engine Optimization (GEO) poisoning. [1] ToE iteratively decomposes a claim, retrieves evidence from multiple sources, and verifies it through an explainable chain. [1] The retrieval process is driven by a reinforcement learning policy, and the authors provide a theoretical analysis that derives a formal error bound, guaranteeing the learned policy converges to a neighborhood of the information-theoretically optimal policy. [1] In experiments across multiple datasets and backbone LLMs, ToE achieved improvements ranging from 4 to 24 percentage points over competitive baselines. [1] The gains were particularly pronounced on adversarially poisoned inputs, a scenario the authors identify as a growing threat to information ecosystems. [1] The paper's abstract notes that the rapid spread of fake news is exacerbated as AI-generated misinformation under GEO poisoning allows crafted content to be systematically surfaced by retrieval systems. [1] The concept of using agent-based optimization to navigate a complex search space draws on established principles in computer science. Ant colony optimization, a probabilistic technique for finding good paths through graphs, uses artificial agents that record their positions and the quality of their solutions to guide later iterations toward better outcomes. [3] This family of algorithms, part of the broader field of swarm intelligence, was initially proposed by Marco Dorigo in 1992 and has since been applied to problems such as vehicle routing and internet routing. [3] ToE applies a similar agent-driven logic to the problem of evidence retrieval for fact-checking. The paper appears on arXiv, a widely-recognized open-access archive for scholarly articles in fields including computer science. [7] The platform has integrated with Hugging Face Spaces to allow authors and the community to link interactive demos directly to a paper's abstract page, enabling readers to try models without writing code. [5] This integration, launched in collaboration with arXivLabs, is designed to increase the reproducibility of research and allow a wider audience to explore computational models. [5]