Steering Vectors

Steering vectors are directions in activation space that capture the difference between two classes of inputs (e.g., positive vs negative sentiment). Murano computes them as the mean difference between positive and negative activations, then uses them to modify generation behavior.

Quick API

import murano

model = murano.Model("meta-llama/Llama-3.2-1B-Instruct")

direction = model.find_direction(
    positive=["What a wonderful, delightful day", "This is fantastic and uplifting"],
    negative=["What a miserable, dreadful day", "This is awful and depressing"],
    layers=[10, 11, 12, 13],
)

print(direction.best_layer)         # layer with highest separation
print(direction.separation_scores)  # {layer: score} dict

Then use the direction to steer generation:

# Add the direction (amplify the concept)
steered = model.generate("Tell me about", steer=(direction, 1.5))

# Ablate the direction (remove the concept)
ablated = model.generate("The movie was", ablate=direction)

The steer parameter takes a (direction, alpha) tuple. Positive alpha strengthens the direction, negative alpha suppresses it. ablate projects the direction out of the residual stream entirely.

Pipeline API

For structured experiments, use the SteeringVector step:

from murano import MuranoDataset, MuranoModel, Pipeline
from murano.steps import Load, Record, SteeringVector

model = MuranoModel("meta-llama/Llama-3.2-1B-Instruct")

dataset = MuranoDataset.contrastive(
    positive=["What a wonderful, delightful day", "This is fantastic and uplifting"],
    negative=["What a miserable, dreadful day", "This is awful and depressing"],
    template_fn=model.chat_template,
)

results = Pipeline([
    Load(dataset),
    Record(model, layers="all", position="last"),
    SteeringVector(normalize=True),
]).run()

steering = results["steering"]
print(steering.best_layer)
print(steering.direction_per_layer[steering.best_layer].shape)  # [d_model]

How it works

SteeringVector computes the contrastive mean difference for each layer:

1. Takes the ActivationStore from the Record step
2. For each layer: direction = mean(positive) - mean(negative)
3. Optionally normalizes each direction to unit norm
4. Computes a separation score (normalized distance between projected class means)
5. Reports the best_layer with the highest separation

Saving and loading directions

from murano import save_results, load_steering

# Save everything
results.save(output_dir="my_experiment", model_id=model.model_id)

# Later, load just the steering result
steering = load_steering("my_experiment/direction/steering.pt")
ablated = model.generate("The movie was", ablate=steering)