Scale every load-bearing finding across Qwen-2.5 sizes (0.5B → 32B)

Goal

Re-run the load-bearing experiments from the past two months across multiple Qwen-2.5 sizes to characterize which findings are size-invariant, which strengthen or weaken with scale, and which only appear above some threshold.

Motivation

Everything in this project so far is Qwen-2.5-7B. Sibling papers (Persona Vectors; Persona Features Control EM) used GPT-4o, so our scaling story is empty. Before the paper, we need to know which of our findings are 7B-specific and which generalize.

Predecessor: #9 (April) covered persona separability + EM susceptibility scaling. This task is the broader sweep over everything that's landed since.

Candidate findings to re-test at scale

(Subset — to be triaged before launch; not all are cheap.)

• Marker leakage / ※-marker dynamics — clean-results #260, #354, #281, #354, #395, #396. Does the single-token ※ log-prob trajectory have a scale threshold? Does the EOS-masked leakage pattern hold below 7B?
• Trait transfer + persona controllability — does the persona-coupling structure observed at 7B emerge sharper at 14B/32B or wash out at 1.5B/3B?
• Belief vs retrieval (contrastive SFT) — #381 / #389. Does the gating effect (predicate emission gated by persona) scale, or is it a 7B capacity artifact?
• Persona spread / persona-pair geometry — #192, #281, #369. Does the cosine-geometry structure require scale?
• Dose-response (EM induction steps) — #139 cliff at 10-25 steps. Does the cliff shift with size?
• Composition / multi-persona — #390. Does additive persona composition hold below 7B?

Scope notes

• Qwen-2.5 family: 0.5B / 1.5B / 3B / 7B / 14B / 32B (instruct). 72B optional if signal.
• Tiered rollout: smallest 4 first (cheap), add 14/32B where the 0.5→7B trend warrants it.
• Many of these need their own follow-up task once we decide which to fold in vs spawn separately.
• Compare-against #9 — may want to merge or supersede it.

Open questions before planning

• Which subset is must-have for the paper vs nice-to-have?
• Should each finding be its own scaling task, or one mega-sweep?
• Budget cap — full sweep across 6 sizes × 6 findings × seeds is hundreds of GPU-hours.