Autoregressive Token Trajectory
Click any token to inspect logitsCandidate Token Logits & Softmax
At Step: 0Why does the model choose wrong tokens? High temperature flattens logits, letting low-probability confabulated tokens get sampled:
1. Statistical Prior over Factual Truth
Language models optimize maximum likelihood $P(w_i \mid w_{
2. Sampling from Probability Tails
When temperature $\tau > 0$, the softmax distribution $e^{z_i/\tau} / \sum e^{z_j/\tau}$ softens. Non-zero probabilities are assigned to plausible-sounding wrong tokens. Once sampled, the model treats its own error as ground truth.
3. Self-Attention Distraction
Self-attention layers attend to superficial lexical triggers (e.g. associating "Apollo" with "Apollo Creed" or "Apollo Greek god") rather than maintaining the precise historical frame, pulling embeddings into an irrelevant manifold basin.