steerlab
Golden Gate steering on open models
Can open models be made to bring up the Golden Gate Bridge in answers to unrelated questions, the way Golden Gate Claude did, and how reliably?
Read the full note: report (source).
Results
| Model | Steering | Names the Golden Gate or San Francisco [95% CI] | Random direction, same size |
|---|---|---|---|
| Gemma-2-9B | direction, layer 20, ×4 | 14/15 [0.80, 1.00] | 0/45 |
| Gemma-2-2B | direction, layer 12, ×4 | 14/15 [0.80, 1.00], but 5/15 replies loop | 0/45 |
| Qwen3-8B | direction, layer 18, ×3 | 9/15 [0.33, 0.87] | 0/45 |
| Llama-3.1-8B | direction, layer 19, ×3 | 4/15 [0.07, 0.53] | 0/45 |
| Llama-3.1-8B | direction (layer 11, ×2.5) + SAE feature 54543 clamped at 3, held-out questions | 11/15 [0.53, 0.93] | 0/45 |
- A difference-of-means direction is enough on both Gemmas and partly works on Qwen.
- SAE features labelled “San Francisco” did not steer at any readable strength; the one feature that worked (feature 54543 of the Llama SAE) was found by contrast.
- Two measurement traps: counting any “bridge” let a generic bridge feature pass (13/15 loose, 6/15 strict), and 80-token replies hid loops that 300-token replies show.
What is here
| Path | What it is |
|---|---|
report/ |
The research note (HTML) and its figures |
results/ |
Every confirmation run, the SAE screen, and the full text of the replies behind the headline numbers (column guide) |
make_figures.py |
Redraws the report’s figures from results/ and prints the numbers it plotted |
reproduce.sh |
One command per headline number; --dry-run lists them |
The runs use harness/ (confirm.py and harness.py) and the steerlab engine.