LAP Steering Demo — Entity Redirect
Paper: Linear Accessibility Profile: Predicting Steering Vector Effectiveness
The key metric, A_lin, applies the model's own unembedding matrix to intermediate hidden states — one forward pass, no training. It measures how "output-aligned" a concept is at each layer and predicts which layer to steer at.
This demo: steer London-answer prompts toward a target city of your choice, using OLMo-2-1B-Instruct.
- Baseline — unsteered model output (answers "London")
- LAP-recommended layer — steered at the A_lin peak → answers with the target city
- Middle layer (standard practitioner heuristic) — steered but A_lin ≈ 0 → no effect
The bar chart shows A_lin across all layers. Steering has no effect where A_lin = 0 — regardless of strength. Only once A_lin rises does intervention redirect the model's output.
Steering quality varies by target. Paris is the clearest example (A_lin = 0.45 at the peak layer, matching the paper). Other cities have weaker A_lin signals and may steer less reliably — this is precisely what the metric predicts. You can type any prompt, but London-framed questions show the effect most cleanly; arbitrary prompts may produce unexpected completions.
Note: This reproduces the entity-steering experiment from §5 / Appendix B of the paper. α = 1.0 matches the paper's setting. A_lin uses the raw logit lens (no tuned-lens correction).
Example London-answer prompts (click to load):
—
Outputs
Unsteered baseline
LAP-recommended layer
Middle layer (standard heuristic)