xu-2024-three-stage-cause-mapping

IN premise — summaries/2026/08/24/xu-2024-hallucination-innate-s6-conclusion.md

Created 2026-08-24T17:11:30+00:00

Xu et al. 2024's three-stage causal model maps specific hallucination causes to stages: data (long-tail knowledge gaps, bias, misinformation), training (exposure bias, diluted attention, architectural deficiencies), and inference (softmax bottleneck, sampling randomness, insufficient context attention).

Summary

This gives a structured diagnosis of why language models produce wrong or fabricated answers, sorting the causes into three clear buckets: gaps and biases in the training data, problems baked in during the training process itself, and limitations in how the model generates responses at query time. The practical upshot is that instead of treating "hallucination" as one vague failure, teams can pinpoint which stage is responsible and target fixes accordingly.