---

name: pplm_perturbation_loop description: Run PPLM-style iterative normalized gradient perturbations with KL regularization and optimizer trace evidence.

PPLM Perturbation Loop

Use this skill when a recovery or implementation needs the central PPLM mechanism: optimizing an inference-time perturbation while the base language-model parameters remain fixed.

Inputs

• Base logits or hidden-state proxy values.
• Objective callback returning loss, gradient, and diagnostics.
• Step size, iteration count, optional KL regularization scale, and gradient normalization flag.

Outputs

• Perturbed logits/proxy values.
• Per-step trace with loss, target mass, KL to the base distribution, and parameter values.
• Booleans showing an optimizer step changed the perturbation.

Workflow

1. Initialize perturbation to zeros or supplied values.
2. Evaluate the attribute objective on base + perturbation.
3. Add a KL-style gradient term toward the base distribution when requested.
4. Normalize nonzero gradients, take a descent step, and log before/after values.
5. Return the final perturbed logits and trace.

Validation

Run python tests/test_perturbation_loop.py or use the Distiller validator.

Limitations

The helper is a faithful low-dimensional proxy for PPLM optimization, not a transformer past-key-value implementation.