Everyone claims their model is grounded, but the data tells a different story....

https://wiki-net.win/index.php/Why_TruthfulQA_and_HaluEval_Are_Misleading_for_2025-2026_Models

Everyone claims their model is grounded, but the data tells a different story. In 2026, hallucination rates swing wildly depending on which benchmark you trust. For example, HalluHard still shows a 30.2% error rate even with live web search enabled