An interactive transformer autoregression visualizer exposing attention dispersion, prior frequency bias, and latent trajectory drift.
Softmax probabilities $\sigma(z_i / T)$ decomposed into Factual Grounding vs Prior Frequency:
Prior training co-occurrence (β = 0.65) overwhelmed context grounding (α = 0.40). At step 2, the model sampled a plausible statistical collocate rather than retrieving precise factual weights.
Language models are trained via cross-entropy loss to predict the next token: $P(w_t \mid w_{1:t-1})$. They optimize for linguistic plausibility and corpus statistical frequency, not an immutable world-state knowledge graph. "Plausible sounding" sequences often dominate sparse factual truths.
Generation is inherently Markovian with respect to context. If a model drifts by even $\epsilon$ at step $k$, that hallucinated token $w_k$ becomes ground truth in its own context window for step $k+1$, triggering an exponential trajectory divergence known as exposure bias.
High context lengths or dense distractors dilute attention weights $\text{softmax}(QK^T / \sqrt{d_k})$. Linear representations in polysemantic neurons cause conceptual bleed, leading the model to conflate entities with similar embedding vectors.