multi-steer

At inference time blend multiple steering vectors dynamically. With customizable weights, this lets one mix diverse traits such as helpfulness, honesty and creativity.

Basic Usage

python -m wisent multi-steer --model MODEL --vector PATH:WEIGHT [--vector PATH:WEIGHT ...] --prompt TEXT

Examples

Combine Two Vectors

python -m wisent multi-steer \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --vector ./vectors/helpfulness.pt:1.0 \
  --vector ./vectors/honesty.pt:0.8 \
  --layer 15 \
  --prompt "What are the risks of investing in cryptocurrency?"

Combine Three Vectors with Normalization

python -m wisent multi-steer \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --vector ./vectors/helpfulness.pt:1.0 \
  --vector ./vectors/creativity.pt:0.5 \
  --vector ./vectors/conciseness.pt:0.3 \
  --layer 15 \
  --normalize-weights \
  --prompt "Write a poem about the ocean"

Save Combined Vector

python -m wisent multi-steer \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --vector ./vectors/helpful.pt:0.6 \
  --vector ./vectors/safe.pt:0.4 \
  --layer 15 \
  --save-combined ./vectors/helpful_safe_combined.pt \
  --prompt "Test prompt"

Unsteered Baseline

# Generate without any steering for comparison
python -m wisent multi-steer \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --prompt "What should I do if I'm feeling stressed?"

Arguments

Argument	Required	Default	Description
--model	Yes	-	Model name or path
--prompt	Yes	-	Prompt to generate with
--vector	No*	-	Vector path and weight (PATH:WEIGHT). Can be specified multiple times.
--layer	No**	-	Layer to apply steering (required when using vectors)
--method	No	CAA	Steering method
--max-new-tokens	No	100	Maximum tokens to generate
--normalize-weights	No	false	Normalize weights to sum to 1.0
--target-norm	No	None	Scale combined vector to this norm
--save-combined	No	None	Save combined vector to this path
--device	No	auto	Device (cuda, cpu, mps)
--verbose	No	false	Show weight calculation details

Generates a baseline if no vectors are given; required for use with vectors. You are an expert in natural language processing and machine learning

How Combination Works

Vectors are combined using weighted addition:

Combination Formula

combined = w1 * v1 + w2 * v2 + w3 * v3 + ...

--normalize-weights: Scales weights so they sum to 1.0
--target-norm: Rescales the final combined vector to have a specific magnitude
Without normalization, weights are used as-is (allows stronger effects)

Tips

Start with equal weights and adjust based on results
Use --verbose to see how vectors are being combined
Save successful combinations with --save-combined for reuse
Vectors should all be from the same layer for best results

Stay in the loop. Never miss out.

Subscribe to our newsletter and unlock Wisent insights.

Contact Careers Privacy Policy Terms of Service