Replicating Validation-Selected Weight Interpolation on SmolLM2-360M Base and Instruct: ARC, HellaSwag and WinoGrande
Year: 2026
Type: preprint
Abstract: Selecting an interpolation coefficient between a base language model and its instruct sibling is often presented as a free improvement over both. We replicate a pre-registered test of this claim on SmolLM2-360M and its instruct version, using the same protocol as for Qwen2.5-0.5B with a coarser six-point grid, selection on 300 validation items per task, and 3,967 disjoint test items. The selected t^*=0.2 scores 55.86% on test against 55.40% for the base model: +0.46 pp (95% paired bootstrap CI [-0.10,+1.02], p=0.113). The interval contains zero, so the hypothesis is not supported. Unlike the Qwen pair, the instruct model is worse than the base model on these log-likelihood benchmarks (no chat template), by -5.64 pp (95% CI [-6.89,-4.35]). Merge accuracy declines from t=0.4 to 1.0; t=0.6 is not distinguishable from the base. A small-t merge is statistically indistinguishable from the base model (no equivalence test). Two model pairs and one hardware platform limit what can be concluded.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.
Source: Figshare
Author: Dheiver Francisco Santos
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233603214
