The Challenge: Adversarial Misinformation

As Generative Engine Optimization (GEO) poisoning becomes more prevalent, adversarial content is increasingly surfacing in retrieval systems, contaminating the reasoning processes of Large Language Models (LLMs). Traditional fact-checking often fails against these systematically crafted inputs because they lack the depth to verify complex, multi-faceted claims.

The ToE Framework: Hierarchical Reasoning

Tree of Evidence (ToE) addresses this by modeling claims not as static strings, but as dynamically expanding argument trees. This hierarchical approach allows the system to:

• Decompose Claims: Break down complex assertions into smaller, verifiable sub-claims.
• Multi-source Retrieval: Use a reinforcement learning-driven agent to fetch evidence from diverse, reliable sources, ensuring the model isn't reliant on a single, potentially poisoned, retrieval path.
• Evidence Aggregation: Systematically evaluate and synthesize the retrieved evidence to build an explainable chain of reasoning that supports or refutes the original claim.

Performance and Theoretical Guarantees

ToE provides a robust defense against adversarial inputs, demonstrating performance improvements of 4 to 24 percentage points over existing baselines. Beyond empirical results, the authors provide a formal theoretical analysis of the retrieval process. They derive an error bound that guarantees the reinforcement learning policy converges to a neighborhood of the information-theoretically optimal policy, providing a mathematical foundation for the framework's reliability in high-stakes information verification.