How Long Do Model Patches Last? A Temporal Perspective on PortLLM

ICLR 2026 Conference Submission22489 Authors

Keywords: LLM, Portability

Abstract: As large language models (LLMs) undergo regular updates through continual pretraining, the temporal reliability of downstream fine-tuning methods becomes increasingly important. Parameter-efficient methods, such as low-rank adaptation (LoRA), offer scalable solutions for task adaptations without requiring full LLM retraining. More recently, PortLLM has been proposed as a training-free patching mechanism that permits patch reuse over consecutive LLM releases. Although these training-free methods are appealing when full fine-tuning is impractical, their temporal reliability remains underexplored. Using PortLLM-style patches as a baseline approach, we conduct large-scale experiments and found that PortLLM patching exhibits a statistically significant performance decline over time, even when the task and neural architecture remain unchanged. Our findings reveal that patch performance degradation is a general and measurable risk when PortLLM is applied over an extended period. The statistical observation of the declining performance trends forms the foundation for our proposed forecasting algorithms, which estimate failure dates and test hypotheses about target-date performance failures. These forecasting algorithms rely on historical performance indicators without requiring downstream fine-tuning or access to original training data. Our framework enables downstream developers to anticipate failure and make informed decisions about when retraining is necessary, thereby supporting reliable and cost-effective LLM maintenance.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 22489