Researchers have created a benchmark called UpgradeBench to evaluate the process of upgrading fine-tuned language models. The benchmark tests how well adapters for specific tasks transfer from one model version to another, and whether it's better to retrain or copy existing adapters. The study found that upgrade gains vary depending on the task and release interval, with some adapters performing worse after a few releases. However, direct adapter copying can be effective, especially when using smaller models. The authors also propose a lightweight method for predicting cross-version adapter portability. --- Why it matters: This research matters to AI engineers because it provides insights into how to efficiently upgrade fine-tuned language models, which are widely used in applications such as chatbots and virtual assistants. By understanding the trade-offs between retraining and copying adapters, developers can make informed decisions about how to maintain their models over time. Source: https://arxiv.org/abs/2608.20918