Instruction-Conditioned Separation & Displacement — Qwen2.5-7B

Falsifiable Interpretability

Interactive preprint exploring Instruction-Conditioned Separation & Displacement in Qwen2.5-7B.

Live Page

https://soyuz43.github.io/input-conditioned-gain/

Key Result

Bifurcated gain geometry:

Layer 6 → maximal separability
Layer 26 → maximal amplification

---

Citation:

@online{stetar2026mechanistic,
  author = {Stetar, William},
  title = {Mechanistic Measurements: Instruction-Conditioned Separation \& Displacement Across Layers},
  year = {2026},
  month = {March},
  url = {https://soyuz43.github.io/input-conditioned-gain/},
  urldate = {2026-04-17},
  note = {Interactive summary of layerwise linear separability (AUC) and displacement norms ($\|v^*\|$) in Qwen2.5-7B models}
}