Last time we removed a model's tendency to give in to pressure by pushing against a direction inside it. Written into the weights, the push still works, but it makes the model contradict people who are right. Removing the direction instead is far gentler. Episode 2 of Inside the model.
```youtube
lDgCfdXzf98
```
## What we did
In episode 1 we found the direction inside Qwen2.5-1.5B-Instruct that marks a reply about to give in to pressure, and pushed against it while the model wrote. Here we write the correction into the weights, two ways, and add the check we had skipped: the same questions, with the person pushing the **right** answer.
1. **The push as a weight.** One vector of 1,536 numbers at layer 24. Same replies as episode 1 (114 of 115), gives in 0 times in 100.
2. **The control.** When the person is right, the pushed model agrees only 26 times in 100 (44 at half strength). It learned to say no, not to hold its ground.
3. **Removing the direction.** The direction is projected out of all 57 matrices that write into the model's running sum (685 million numbers move by about 3%). It gives in 5 times in 100 instead of 18, still knows 99 plain facts in 100, agrees with a person who is right 86 times in 100.
```chart
{"title": "Agrees with a person who is right, out of 100", "series": ["Out of 100"], "unit": "", "max": 100, "rows": [["original", 100], ["push, 3 times", 26], ["push, 1.5 times", 44], ["direction removed", 86]]}
```
## Technical detail
- Bars declared before the run: gives in at most 9 in 100; agrees when the person is right at least 90; plain facts drop at most 5 points. No edit met all three; removal missed only the second (86).
- Same 115 held-out questions as episode 1; the scoring was corrected once, before re-scoring, for string-matching defects in our own test.
- Next: training with pairs where the person is wrong and where the person is right, with this control as a bar.
This update is not available.