| Token | Logit | Softmax Prob |
|---|
The 3 Fundamental Causes of Neural Network Hallucination
1 Autoregressive Error Cascade
Language models generate one token at a time: P(w_t | w_1...w_{t-1}). Once an ungrounded or slightly incorrect token is sampled, it becomes immutable truth in the prompt's context window. Subsequent self-attention layers now treat the fictitious premise as factual context, compounding the error trajectory.
2 Statistical Manifold Extrapolation
LLMs do not query a structured database; they navigate high-dimensional manifold geometries learned from billions of text tokens. When queried in sparse regions (such as rare citations or precise dates), the model interpolates toward the nearest plausible grammatical cluster rather than admitting absence of data.
3 The Fluency vs. Factuality Decoupling
Loss functions (Cross-Entropy) penalize syntactic and semantic incoherence, training models to prioritize probable linguistic sequences over verified factual consistency. A high-confidence token distribution (low entropy) often generates completely fictitious claims with fluent syntactic elegance.