---

name: pplm_fusion_generation description: Fuse PPLM perturbed and unperturbed token distributions using geometric mixing for fluent controlled decoding.

PPLM Perturbed Distribution Fusion

Use this skill after an attribute perturbation has produced controlled logits and the run needs to combine them with the unperturbed language-model distribution.

Inputs

Base logits or probabilities, perturbed logits or probabilities, gm_scale in [0,1], and a deterministic selection mode.

Outputs

A normalized fused distribution, selected token index, entropy, and endpoint diagnostics.

Workflow

1. Convert logits to probabilities when needed.
2. Compute geometric mixture base^(1-gm_scale) * perturbed^gm_scale.
3. Normalize the fused probabilities.
4. Select the argmax token for deterministic recovery tests.

Validation

Run python tests/test_fusion_generation.py or Distiller validation.

Limitations

This skill validates the fusion rule and does not implement full top-k or nucleus sampling.