---

name: pplm_latent_perturbation description: Use when a PPLM-style controller must perturb latent/logit state at inference time with attribute and KL losses while freezing the base model.

PPLM Latent Perturbation

Use this skill to implement the core PPLM inference-time update.

Inputs

• Base next-token logits or latent state.
• Attribute loss function.
• Step size, iteration count, KL scale.

Outputs

• Perturbed probability distribution.
• Trace with before/after loss, KL, and perturbation values.

Workflow

1. Copy the base logits/state and keep it immutable.
2. Initialize a perturbation at zero.
3. Iteratively estimate the gradient of attribute loss plus KL-to-base penalty.
4. Update only the perturbation.
5. Return the final distribution and trace proving the target probability changed.

Validation

Run python tests/test_latent_perturbation.py.

Limitations

Reduced proxy implementations must declare that they are not full GPT-2 345M reproduction.