Alessandro Stolfo’s study evaluates groundedness in retrieval-augmented LLMs focusing on long-form question answering (LFQA). The research assesses whether LLMs are truly grounding each sentence in retrieved documents or if their knowledge stems from prior pre-training data Read More

• Uses 3 datasets and 4 model families to investigate groundedness.
• Uncovers a significant fraction of sentences are ungrounded, despite containing correct answers.
• Contemplates the effects of model size, decoding strategy, and instruction tuning on groundedness.
• Larger models are shown to be more effective in grounding outputs, though substantial hallucinations persist.
• Emphasizes the need for enhanced mechanisms to prevent ungrounded content generation in LLMs.

The study conveys the complexity of ensuring accurate and reliable LLM outputs, highlighting the necessity for developing more sophisticated AI systems that can verify the authenticity of each generated sentence.