Current Prompt Context
"According to the 2021 landmark paper by Henderson et al., the exact rate of..."
Autoregressive Generation Sequence
Directly manipulate probability manifold curvature, attention drift, and entropy decay to inspect logit collapse.
Transformers calculate next tokens via softmax(logits / T). In unfamiliar territory (Out-Of-Distribution), cross-attention keys become diffuse. When all candidate tokens hold low unnormalized energies, softmax artificially amplifies tiny stochastic variances into confident pseudo-facts.
Every generated token is appended directly to the input window: x_(t+1) ~ P(x | x_1...x_t). A single speculative token chosen at step t conditions step t+1 to validate that false assertion, locking the model into a self-reinforcing confabulation spiral.
Ground truth is not an inherent property of neural representations; it is a sparse, constrained manifold inside an astronomical dimensional space. Pushing vectors slightly past the manifold boundary yields syntactically flawless text with complete factual divergence.