jb1999/lap-steering-demo

LAP Steering Demo — Entity Redirect

Interactive demo for the paper Linear Accessibility Profile: Predicting Steering Vector Effectiveness.

The core claim: A_lin (logit lens accuracy at each layer) predicts which layer a steering vector will be effective at — with no training required.

This demo: steer London-answer prompts toward a target city on OLMo-2-1B-Instruct.

What you see

Three completions side-by-side for any prompt:

Alin is computed by applying the model's own unembedding matrix to intermediate hidden states (the logit lens). Where Alin = 0, the concept is not yet linearly decodable — steering at that layer does nothing regardless of strength. This is the paper's key finding.

Steering quality varies by target city. Paris has the strongest signal (A_lin = 0.45 at L13), matching the paper exactly. Other cities are weaker (0.05–0.30) and may steer less reliably — which is precisely what the metric predicts.

Technical details

Model: allenai/OLMo-2-0425-1B-Instruct

Steering vectors: Difference-of-means (CAA). For each target city:

sv[l] = mean(city_activations[l]) - mean(London_activations[l])

A_lin (logit lens accuracy):

A_lin(l) = fraction of prompts where argmax(W_U · LayerNorm(h_l)) == target token

Applied to hidden states at each layer using the model's own final norm and LM head. No training, no probes — one forward pass.

Steering injection:

hidden_states[l] += α × steering_vector[l]

Via register_forward_hook. Hook is registered, used for one generation, immediately removed.

Run locally

git clone https://huggingface.co/spaces/YOUR_USERNAME/lap-steering-demo
cd lap-steering-demo
pip install -r requirements.txt

# Precompute steering vectors + A_lin for all cities (~2 min on GPU, ~15 min on CPU)
python precompute.py

python app.py

Links

• —📄 Paper
• —💻 Code

Run locally

Or with uv:

uv sync
uv run python precompute.py
uv run python app.py