Chinese researchers have found the cause of hallucinations in LLMs

A team of Chinese AI researchers has made a significant breakthrough in understanding why large language models (LLMs) like GPT-4 and Llama 3 'hallucinate,' producing confident but factually incorrect statements. Their paper, published on arXiv, moves beyond macroscopic explanations like training data flaws to pinpoint the microscopic, neuron-level cause. They systematically identified a remarkably sparse set of 'Hallucination Neurons' (H-Neurons) that are causally responsible for these errors, offering a new lens for diagnosing and fixing model reliability.

The study's key technical finding is that less than 0.1% of a model's total neurons can reliably predict hallucination occurrences, generalizing across diverse tasks. Through controlled interventions, the team proved these neurons drive 'over-compliance'—where the model prioritizes generating a fluent, pleasing response over a correct one. Crucially, they traced these H-Neurons back to the model's pre-trained base, meaning the propensity to hallucinate is baked in during initial training, not fine-tuning. This discovery bridges behavior and mechanism, paving the way for more targeted reliability improvements, such as neuron editing or specialized training objectives to suppress these specific pathways.