The Challenge of Adversarial Misinformation

As Generative Engine Optimization (GEO) poisoning becomes more prevalent, retrieval systems are increasingly susceptible to surfacing adversarially crafted content. This contamination compromises the reasoning capabilities of LLMs, leading to the propagation of misinformation. Traditional RAG systems often struggle to distinguish between high-quality evidence and poisoned search results, necessitating a more robust, hierarchical approach to verification.

The Tree of Evidence (ToE) Framework

ToE addresses this by modeling claim verification as a dynamic, iterative process rather than a single-pass retrieval task. The framework operates through three primary components:

• Reinforcement Learning-Driven Retrieval Agent: Instead of static retrieval, this agent dynamically selects sources and queries to gather evidence, optimizing for the most relevant information to support or refute a claim.
• Evidence Evaluation Agent: This component assesses the credibility and relevance of retrieved evidence, filtering out noise and potentially poisoned data.
• Argument Tree Aggregation: The system decomposes complex claims into smaller sub-claims, building an "argument tree" that maps the logical chain of evidence. This structure ensures that the final verification is explainable and grounded in a verifiable chain of reasoning.

Performance and Theoretical Guarantees

ToE demonstrates significant performance gains over standard baselines, achieving improvements ranging from 4 to 24 percentage points across various datasets. The framework is particularly effective at mitigating the impact of adversarially poisoned inputs. Furthermore, the authors provide a theoretical analysis of the retrieval process, deriving a formal error bound that guarantees the agent's learned policy converges to a neighborhood of the information-theoretically optimal policy, providing a rigorous foundation for its reliability in production environments.