A Pipeline to Assess Merging Methods via Behavior and Internals
Yutaro Sigrist, Andreas Waldis · 2025
Read A Pipeline to Assess Merging Methods via Behavior and Internals by Yutaro Sigrist, Andreas Waldis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-19476.txt.
Description
Merging the weights of two trained language models can blend their skills, such as math and code. Yutaro Sigrist and Andreas Waldis ask whether merged models should be judged by more than their task scores. They built a test pipeline that checks both outward behavior and the language facts stored inside the model. Their test case merges chat-tuned, math-tuned, and code-tuned models from one model family.
Merged models usually scored between their two parents on tasks. Yet their stored sense of word forms and sentence shape could beat both parents. Task scores and inner scores also ranked the merges in near-unrelated orders. The authors argue that score-only checks can hide real gains and losses. Their pipeline gives builders a fuller way to judge a merge before use.
Catalog record
| CREATOR | Yutaro Sigrist, Andreas Waldis |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-19476