ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation
06:00 · June 29, 2026 · arXiv cs.AI RSS
The rapid spread of fake news poses increasing threats to information ecosystems, especially as AI-generated misinformation under Generative Engine Optimization (GEO) poisoning allows adversarially crafted content to be systematically surfaced by retrieval systems, contaminating LLM reasoning. In this paper, we propose Tree of Evidence (ToE), a hierarchical evidence reasoning framework for automated fact-checking that models each claim as a dynamically expanding argument tree. ToE integrates a reinforcement learning-driven multi-source retrieval agent, an evidence evaluation agent, and an argument tree aggregation algorithm to iteratively decompose, retrieve, and verify claims through an explainable evidence chain. We further provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the learned policy converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple datasets and backbone LLMs demonstrate that ToE achieves improvements ranging from 4 to 24 percentage points over competitive baselines, with particularly pronounced gains on adversarially poisoned inputs.
Summary
The Tree of Evidence (ToE) framework addresses the growing problem of AI-generated misinformation that exploits Generative Engine Optimization to rank higher in retrieval systems and contaminate LLM outputs. It treats each input claim as the root of a dynamically expanding argument tree and verifies it through iterative decomposition along dimensions such as who, what, when, where, why, and how. Three cooperating agents drive the process: a reinforcement-learning retrieval agent that selects queries and sources, an evidence-evaluation agent that assigns veracity and reliability scores to extracted snippets, and an aggregation algorithm that propagates node scores upward until the root reaches a convergence threshold.
Evidence is gathered from heterogeneous sources including Wikipedia, arXiv, fact-checking sites, search engines, and social media. The retrieval agent is trained as a Partially Observable Markov Decision Process and stops autonomously once additional results are unlikely to improve the judgment. To reduce confirmation bias, the system generates balanced queries for background information, supporting evidence, and counter-evidence at each node. When current evidence proves insufficient, the tree expands new sub-claims, producing an explicit reasoning trace that explains the final veracity score.
A formal error bound derived for the learned retrieval policy shows that it converges to a neighborhood of the information-theoretically optimal policy. Experiments across multiple public datasets and several backbone LLMs report absolute gains of 4 to 24 percentage points over competitive baselines, with the largest margins observed on adversarially poisoned inputs. The authors also release the AdvFact dataset to support further robustness testing under GEO-style attacks.
Why it matters
This research is highly relevant to the Dutch AI market's focus on ethical, transparent, and trustworthy AI. The proposed explainable claim verification framework offers advanced methodologies for researchers and enterprises to combat misinformation and align with EU regulations on AI transparency.
