Hallucination in Language Models

The field studies when language models produce claims that are false, unsupported by their sources, or inconsistent with images, audio, and other evidence. Researchers evaluate errors at the level of claims, spans, reasoning steps, and agent trajectories, using factual question answering, retrieval-augmented generation, summarization, and multimodal benchmarks. Common approaches include retrieval and external verification, confidence calibration and abstention, targeted fine-tuning, and inference-time interventions based on decoding or internal model signals. A central tension is that uncertainty signals and mitigations vary across models and tasks: reducing unsupported answers can also suppress correct ones, while plausible reasoning or fluent outputs do not guarantee factual grounding.