Candidate Logits Distribution H=0.0
Joint Sequence Probability $\prod P(t_i)$ Decay
Autoregressive Compounding
Hallucinations are not random bugs: as the model emits each low-confidence token, errors feed into future context, exponentially decaying factual joint probability.
Softmax Flattening
Elevated temperature flattens token logits ($e^{z_i/T}$), pulling improbable latent noise into high-probability selection bounds.
Grounding Attractors
Context grounding (RAG) acts as an attention anchor, penalizing non-factual attractor basins and suppressing phantom citations.