From Trust to Verification: Teaching AI How to Think, Not What to Believe

Modern research agents excel at retrieving and summarizing information but falter when the truth is contested. My work rethinks how large language models are trained to reason, introducing IMMUNE-Bench, a framework that teaches AI systems to navigate misinformation through verification rather than trust.

IMMUNE-Bench builds synthetic, evolving worlds—complete with conflicting claims, temporal shifts, and probabilistically modeled biases. Within these worlds, agents learn to weigh evidence, cross-reference sources, and calibrate their confidence instead of parroting dominant narratives. The system’s three-tier architecture orchestrates document generation, summarization, and temporal evolution, enabling thousands of interconnected events to unfold coherently over simulated time.

A reinforcement learning framework drives the training process, rewarding factual accuracy, synthesis quality, and epistemic humility. Agents are incentivized to compare sources, recognize fabrication, and abstain when evidence is inconclusive. This emphasis on how conclusions are reached—rather than simply what is correct—cultivates genuine reasoning resilience.

By moving beyond clean, consensus-driven datasets, this research aims to create AI systems capable of critical inquiry—models that can reason under uncertainty, question their own assumptions, and remain grounded in verification even when truth itself becomes adversarial.