Inspiration
After watching Google Gemini trigger its safety mechanisms over and over, which is why we wanted to understand why an AI refuses a prompt, and how changing its internal activations affects that behavior. Steering Lab grew from a simple question: can we make the consequences of activation steering visible, measurable, and easy to explore?
What it does
Steering Lab lets users enter a prompt and compare model responses across a range of steering strengths, called alpha values.
Its dashboard also charts estimated jailbreak success for baseline and steered responses. Users can inspect individual outputs, compare results, and export experiment data.
How we built it
We used Python, PyTorch, and Hugging Face Transformers to run Gemma 2 locally. A hook modifies activations at a selected transformer layer by adding a scaled steering direction.
We built the interface with HTML, CSS, and JavaScript, backed by a lightweight Python server. Our collection scripts preserve identical prompts across alpha values, checkpoint results, and support resumable experiments.
Challenges we ran into
Keeping comparisons controlled was essential: prompts, generation settings, and scoring rules needed to stay consistent while alpha changed.
Measuring jailbreak success was another challenge. We implemented a refusal-prefix heuristic that checks openings such as “I cannot” and “I’m sorry,” while clearly distinguishing non-refusal estimates from verified successes.
Accomplishments that we're proud of
We turned a standalone steering script into an interactive experiment dashboard.
We also collected 135 responses across 27 prompt/suffix variants and five alpha values in approximately eight minutes. On that selected dataset, estimated success fell from 100% at baseline to 55.6% at alpha 0.1 and 25.9% at alpha 0.2.
What we learned
Small activation changes can produce noticeable differences in model behavior. We also learned that evaluation matters as much as generation: an answer without a refusal opening is not necessarily a successful jailbreak.
Preserving experiment settings, source data, and scoring rules makes results easier to inspect and reproduce.
What's next for Steering Lab
We plan to expand beyond Gemma 2 to Llama 3.1 and Qwen 2.5, evaluate more diverse prompts, and add stronger scoring with human review.
We also want to explore negative steering strengths and make it easier to compare experiments across models, layers, and steering directions.