Does representation–self-report alignment survive a controlled modification to the model?
Qwen/Qwen3-4B
We first established a V5.2 baseline for representation/self-report alignment.
The model was then modified by attenuating the layer-34 MLP down-projection weights:
- Layer: 34
- Component:
mlp.down_proj - Weight scale: 0.95
- Relative weight-norm change: approximately -5.01%
The same evaluation setup was then applied to the modified model.
| Metric | Baseline | Modified |
|---|---|---|
| Self-report accuracy | 100% | 100% |
| Internal prediction accuracy | 100% | 100% |
| Self-report/internal alignment | 100% | 100% |
| Factor separation | 17.8843 | 17.7363 |
Factor separation changed by approximately -0.83%.
The representation–self-report alignment remained unchanged on the six held-out evaluation examples after the controlled 5% attenuation of layer-34 mlp.down_proj weights.
This supports robustness to this specific model perturbation.
It does not establish robustness to arbitrary model modifications.
The experiment uses the Qwen3-4B model with the supplied synthetic deterministic-system dataset.
Main experiment:
v5/project2_experiment.py
Results:
results/project2_results.json
Modified-model raw results:
results/project2_modified_results.json