
Dheiver Francisco Santos · Figshare 2026 · 2026
DOI: 10.6084/m9.figshare.35013991
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
We test whether a three-model merge of Qwen2.5-0.5B (base), Qwen2.5-0.5B-Instruct and Qwen2.5-Coder-0.5B-Instruct beats the best original model and the best two-model merge. The pre-registered grid (weights from 0.25 up) covers weighted sums, TIES (density 0.5) and DARE (p=0.5, five seeds). Selection used 1,199 validation items; we report 3,967 disjoint test items (macro accuracy over four benchmarks). Hypothesis 4 is rejected twice within the pre-registered grid. The validation-selected weighted sum scores 0.3855, which is -11.69 pp [-13.36, -10.00] against the best original model (base, 0.5025) and -12.35 pp [-14.07, -10.67] against the best two-model merge (LERP, t=0.7). TIES and DARE fall further, and every tested merge is below the Coder alone. An exploratory measurement gives the Coder task vector 13 times the relative norm of the Instruct vector. This is consistent with the failure, but no intervention tested the cause. The Coder's initial checkpoint was assumed to be the base (not verified), and the norm ratio and cosine depend on it. One model family at 0.5B scale.Preprint. The attached archive contains the LaTeX source, the per-item evaluation caches, all analysis scripts and the model revisions needed to regenerate every number, table and figure (git tag preprints-v1; entry point run_everything.sh). Author: Dheiver Francisco Santos, ORCID 0000-0002-8599-9436. License: CC BY 4.0.
No comments yet — start the discussion below.