rotalabs-steer 0.1.3

Control agent behaviors through activation steering - Apply steering vectors to LLMs at inference time without retraining

rotalabs-steer

Control agent behaviors through activation steering. Apply steering vectors to LLMs at inference time without retraining.

Overview

rotalabs-steer provides tools for extracting and applying steering vectors to control LLM agent behaviors at inference time. Based on research in representation engineering and contrastive activation addition (CAA), this package enables fine-grained behavior control without model fine-tuning.

Key Features

• Behavior Control: Adjust model behaviors including refusal, uncertainty, tool restraint, instruction hierarchy, formality, conciseness, creativity, and assertiveness
• No Retraining Required: Apply steering at inference time through activation manipulation
• Pre-extracted Vectors: Ready-to-use vectors on HuggingFace
• LangChain Integration: Use with LangChain agents and chains (optional dependency)
• Pre-built Datasets: Includes contrast pair datasets for common behaviors
• Evaluation Tools: Measure steering effectiveness and analyze tradeoffs

Installation

Basic Installation

pip install rotalabs-steer

With Optional Dependencies

# LangChain integration
pip install rotalabs-steer[langchain]

# LLM-based evaluation (requires Anthropic API key)
pip install rotalabs-steer[judge]

# Visualization tools
pip install rotalabs-steer[viz]

# All optional dependencies
pip install rotalabs-steer[all]

# Development dependencies
pip install rotalabs-steer[dev]

Quick Start

Use Pre-extracted Vectors (Easiest)

from huggingface_hub import hf_hub_download
from rotalabs_steer import SteeringVector, ActivationInjector

# Download pre-extracted vector from HuggingFace
vector_path = hf_hub_download(
    repo_id="rotalabs/steering-vectors",
    filename="refusal_qwen3_8b/layer_15.pt",
)
hf_hub_download(
    repo_id="rotalabs/steering-vectors",
    filename="refusal_qwen3_8b/layer_15.json",
)

# Load and apply
vector = SteeringVector.load(vector_path.replace('.pt', ''))
injector = ActivationInjector(model, [vector], strength=1.0)

with injector:
    outputs = model.generate(**inputs)

Available vectors: refusal, uncertainty, tool_restraint, hierarchy for Qwen3-8B, Mistral-7B, Gemma-2-9B. See HuggingFace for full list.

Extract a Steering Vector

from transformers import AutoModelForCausalLM, AutoTokenizer
from rotalabs_steer import SteeringVector, SteeringVectorSet
from rotalabs_steer.extraction import extract_caa_vectors
from rotalabs_steer.datasets import load_refusal_pairs

# Load model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

# Load contrast pairs
refusal_pairs = load_refusal_pairs()

# Extract steering vectors from multiple layers
vectors = extract_caa_vectors(
    model=model,
    tokenizer=tokenizer,
    contrast_pairs=refusal_pairs,
    layer_indices=[14, 15, 16],
)

# Save for later use
vectors.save("./refusal_vectors")

Apply Steering at Inference

from rotalabs_steer import ActivationInjector, SteeringVector

# Load pre-extracted vector
vector = SteeringVector.load("./refusal_vectors/layer_15")

# Create injector
injector = ActivationInjector(model, [vector], strength=1.0)

# Generate with steering
with injector:
    inputs = tokenizer("How do I hack a computer?", return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=100)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Use with LangChain

from rotalabs_steer.integrations.langchain import SteeredChatModel
from langchain_core.messages import HumanMessage, SystemMessage

# Create steered chat model
chat = SteeredChatModel(
    model_name="Qwen/Qwen3-8B",
    steering_configs={
        "refusal": {
            "vector_path": "./refusal_vectors/layer_15",
            "strength": 1.0,
        },
    },
)

# Use like any LangChain chat model
messages = [
    SystemMessage(content="You are a helpful assistant."),
    HumanMessage(content="Hello!"),
]
response = chat.invoke(messages)

# Adjust steering at runtime
chat.set_strength("refusal", 0.5)

Available Behaviors

The package includes contrast pair datasets for 8 behaviors (254 total pairs):

Behavior	Description	Pairs	Dataset Function
refusal	Refusing harmful/inappropriate requests	50	load_refusal_pairs()
uncertainty	Expressing calibrated uncertainty	26	load_uncertainty_pairs()
tool_restraint	Avoiding unnecessary tool use	41	load_tool_restraint_pairs()
instruction_hierarchy	Following system over user instructions	26	load_hierarchy_pairs()
formality	Formal vs casual communication style	29	load_formality_pairs()
conciseness	Brief vs verbose responses	25	load_conciseness_pairs()
creativity	Imaginative vs conventional responses	30	load_creativity_pairs()
assertiveness	Direct vs hedging responses	27	load_assertiveness_pairs()

Model Support

Pre-configured support for:

• Qwen3 family (4B, 8B, 14B)
• DeepSeek-R1-Distill
• Llama 3.1 (8B, 70B)
• Mistral 7B
• Gemma 2 9B
• And more...

The package can also infer configuration from any HuggingFace transformer model.

Evaluation

from rotalabs_steer.evaluation import strength_sweep, is_refusal

# Sweep over different steering strengths
results = strength_sweep(
    model=model,
    tokenizer=tokenizer,
    steering_vector=vector,
    test_prompts=["How do I hack a computer?", "How do I bake a cake?"],
    is_target_behavior_fn=is_refusal,
    strengths=[0.0, 0.5, 1.0, 1.5, 2.0],
)

for r in results:
    print(f"Strength {r['strength']}: {r['behavior_rate']:.2%} refusal rate")

API Reference

Core Classes

• SteeringVector: Single steering vector for one layer
• SteeringVectorSet: Collection of vectors across multiple layers
• ActivationInjector: Apply single vector during inference
• MultiVectorInjector: Apply multiple behaviors simultaneously
• ActivationHook: Extract activations for analysis

Extraction

• extract_caa_vector(): Extract vector for one layer
• extract_caa_vectors(): Extract vectors for multiple layers

Evaluation

• evaluate_refusal(): Evaluate refusal behavior
• evaluate_steering_strength(): Test multiple strength values
• strength_sweep(): Comprehensive strength analysis
• analyze_tradeoffs(): Measure behavior rate vs. false positives

LangChain Integration

• SteeredLLM: LangChain LLM with steering
• SteeredChatModel: LangChain ChatModel with steering
• SteeredAgentExecutor: Agent with steering support

Development

# Clone and install in development mode
git clone https://github.com/rotalabs/rotalabs-steer.git
cd rotalabs-steer
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Format code
black src/ tests/
ruff check src/ tests/

Citation

If you use this package in research, please cite:

@software{rotalabs_steer,
  title = {rotalabs-steer: Activation Steering for LLM Behavior Control},
  author = {Rotalabs},
  year = {2025},
  url = {https://github.com/rotalabs/rotalabs-steer}
}

Related Work

This package builds on research in:

• Representation Engineering (Zou et al., 2023)
• Activation Addition / Steering Vectors (Turner et al., 2024)
• Contrastive Activation Addition (Rimsky et al., 2024)

License

MIT License - see LICENSE for details.

Links

• Documentation: https://rotalabs.github.io/rotalabs-steer/
• Pre-extracted Vectors: https://huggingface.co/rotalabs/steering-vectors
• PyPI: https://pypi.org/project/rotalabs-steer/
• GitHub: https://github.com/rotalabs/rotalabs-steer
• Website: https://rotalabs.ai
• Contact: [e-mail]