Work

SLERP versus LERP for Merging Qwen2.5-0.5B Base and Instruct: A Paired Comparison on ARC, HellaSwag, and WinoGrande

Year: 2026
Type: preprint
Abstract: Spherical linear interpolation (SLERP) is an alternative to linear interpolation (LERP) for merging fine-tuned language models. We merge Qwen2.5-0.5B and its Instruct sibling at four coefficients with both operators and score macro accuracy on 3,967 held-out items from four benchmarks. The pre-registered hypothesis was that SLERP does not differ from LERP. Paired SLERP-LERP test differences range from -0.10 to +0.35 percentage points (pp), and every 95% bootstrap interval contains zero. No difference was detected (no equivalence test). The 95% interval for the exploratory mean difference is [-0.09,+0.39] pp. The test used bfloat16 weights, whose rounding is about ten times larger than the relative SLERP–LERP gap, and about 90% of merged elements are identical after the cast. On validation in fp32, the difference is 0.00 and -0.17 pp at t=0.4 and 0.6 (bf16: +0.25 and +0.42): the size of precision noise, with the sign reversed at t=0.6. The primary comparison, a validation-chosen SLERP merge against the best original model, is not distinguishable (+0.40 pp, p=0.262). One model pair.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
Language: English
Open Access status: green
OpenAlex ID: W7233595927
Cites: 33
Cited by:
Related to: