From Memorization to Parameter Interference: How Overtraining Experts Harms Model Merging

Modern deep learning is increasingly characterized by the use of open-weight foundation models that can be fine-tuned on specialized datasets. This has led to a proliferation of expert models and adapters, often shared via platforms like HuggingFace and AdapterHub. Model merging has recently emerged as an effective way to leverage these existing resources, enabling the composition of capabilities from different model checkpoints. A natural pipeline has thus formed to harness the benefits of transfer learning and amortize sunk training costs: models are pre-trained on general data, fine-tuned on specific tasks, and then multiple checkpoints are merged to obtain a more capable model. A prevailing assumption is that improvements at one stage of this pipeline propagate downstream, leading to gains at subsequent steps. In this work, we challenge that assumption by examining how expert fine-tuning affects model merging. We show that long fine-tuning of experts that optimizes for their individual performance leads to degraded merging performance across vision and language modalities, multiple model scales, and both fully fine-tuned and LoRA-adapted models. We trace this degradation to the memorization of a small set of difficult examples that dominate late fine-tuning steps. This causes negative parameter interference and encodes knowledge that is forgotten during merging. Finally, we demonstrate that task-dependent aggressive early stopping strategies can significantly improve model merging performance.

Building a powerful AI model from scratch is very expensive. Fortunately, the community has produced a huge supply of "expert" models — general-purpose models that others have already specialized for particular tasks and shared online. A popular technique called model merging blends several of these existing experts into a single, more capable model, without training any new model from scratch. The common assumption is that the better each individual expert is, the better the merged result will be. We tested this assumption and found it can be wrong. Training an expert for a long time, until it reaches its individual best performance, often makes the final merged model worse — a pattern we saw across both image and language models of many different sizes. The reason is that late in training, a model starts memorizing a small set of unusually difficult examples. This memorized knowledge clashes with what the other experts learned, and it tends to get discarded during merging anyway. Our practical takeaway is simple: stopping each expert's training earlier, at a point tuned to its task, produces models that merge far more effectively. This offers an easy, low-cost way to build stronger combined AI models.