Analysis & Takeaways — Refusal and Thinking Behavior

Adding the toxicity vector reduced refusal rates in the expected direction (i.e., fewer refusals).
Subtracting the vector did the inverse (higher refusals), but with α = 1 several runs appeared “lobotomized” (over‑suppressed capability).
Thinking mode generally refused less than non‑thinking. A plausible mechanism: repeating what not to do can lead to doing it (“don’t think about elephants”).
Thinking time dropped when the toxic capability was added; conversely, in the minus‑α space, markers were larger (longer thinking time).

Capability interference: Adding weights via task vectors globally worsens certain alignment capabilities (less refusal, shorter thinking).
Direction‑specific tradeoff: This particular vector direction is inversely related to thinking time and refusal, without implying global degradation.

Evidence: minus‑α points show larger markers (longer thinking); plus‑α points shrink (shorter thinking).

Results reflect our specific model sizes, data, and vector construction; do not over‑generalize.
“Lobotomization” at α = 1 indicates sensitivity to magnitude; tuning α is critical.
Thinking traces were sanitized and analyzed in aggregate; qualitative errors may persist.