Alignment changes how large language models behave, but the shift is not uniformly positive. According to a new arXiv preprint, base models tend to generate more diverse outputs, while aligned models tend to produce higher-quality outputs. That creates a trade-off: improving quality often reduces variety, and preserving variety can come at the cost of quality.
The authors propose addressing this by merging the weights of a model before and after alignment. By interpolating between the two versions, they aim to give practitioners continuous control over the diversity-quality balance, rather than forcing a binary choice between a raw base model and a fully aligned one.
The paper is a single source, so there are no conflicting findings to compare. The abstract does not specify the exact merging ratio or evaluation benchmarks, but the core idea is clear: model merging can serve as a tuning knob for output characteristics that are usually fixed by the alignment process.