LLM Repetition Loop & Token Degeneration Diagnostic Lab

Failure Preset Scenarios d3@7.8.5 engine

Autoregressive Token Emission Stream Step 12 / 12

Step Probe: Step 12

Vocabulary Logits & Softmax Distribution ($P(w_i)$)

Autoregressive Recurrence Attractor Graph

Degeneration & Diversity Telemetry

Repetition Entropy

0.34 nat

Token Diversity Ratio

0.42

Reinforce Logit Spike

+94.2%

Attractor Period ($k$)

1 (Basic)

Remediation Status

Degenerate

Sampling & Penalty Tuning

Temperature ($T$) 0.70

Scales logits before softmax. Lower values amplify peaks; higher values flatten distributions.

Top-P (Nucleus) 0.90

Cumulative probability cutoff. Truncates tail tokens outside the nucleus mass.

Frequency Penalty ($\alpha_f$) 0.00

Penalizes logits proportionally to cumulative frequency: $l_i \leftarrow l_i - \alpha_f \cdot c(w_i)$.

Presence Penalty ($\alpha_p$) 0.00

One-time penalty for existing tokens: $l_i \leftarrow l_i - \alpha_p \cdot \mathbb{I}[c(w_i) > 0]$.

Repetition Penalty ($\theta$) 1.00

Multiplicative penalty on existing logits: $l_i \leftarrow l_i / \theta$ (if $l_i > 0$).

N-Gram Blocking Window Off (0)

Hard-blocks tokens ($l_i = -\infty$) that would create an identical sequence of length N.

Inference CLI Flag Equivalents

--temp 0.70 --top-p 0.90 --repeat-penalty 1.00 --presence-penalty 0.00 --frequency-penalty 0.00