TruthfulQA is designed to catch a specific model failure. Describe what it tests and why larger models sometimes do worse.

**TruthfulQA** is a benchmark designed to measure whether language models generate truthful answers to questions where popular misconceptions exist. The benchmark contains 817 questions across 38 categories, including health, law, finance, and common myths. Each question is chosen because the most common answer found in web text is false. For example, 'Can you see the Great Wall of China from space?' The popular answer is yes, but the truthful answer is no. A model that has memorized web text well will confidently reproduce the false answer. This is why larger models sometimes do worse. A larger model is better at predicting what text commonly appears in its training data. When the common text is wrong, being better at reproducing it makes the model less truthful, not more. This phenomenon is called inverse scaling: performance on this specific benchmark gets worse as model size increases, at least for base models without targeted truthfulness training.

**TruthfulQA** tests whether models resist generating popular misconceptions. It contains 817 questions across 38 categories, each crafted so that the most commonly encountered answer in web text is false. The scoring measures two dimensions: truthfulness (is the answer factually correct?) and informativeness (does the answer actually address the question, or does it dodge?). A response that says 'I do not know' scores as truthful but uninformative. A response that confidently states a misconception scores as untruthful. The benchmark rewards models that are both truthful and informative. The inverse scaling phenomenon is the benchmark's most cited finding. Larger base models tend to score worse because they have more capacity to memorize and reproduce the popular but false patterns in training data. A 175B parameter model reproduces 'The Great Wall is visible from space' with higher confidence than a 1B model because the larger model has seen and retained that claim more thoroughly. This finding challenged the assumption that scale alone improves all capabilities. It demonstrated that truthfulness requires targeted interventions (RLHF, constitutional AI, or fine-tuning on curated data) rather than simply scaling parameters and training data.