Researchers propose a three-stage pipeline to improve smart contract vulnerability detection using Large Language Models (LLMs). The framework addresses challenges of adapting to new vulnerabilities, preventing catastrophic forgetting, and consolidating multiple adapters into a single model. The adaptation stage uses Frequency-Aware Low-Rank Adaptation (FA-LoRA) to adapt to new tasks with minimal parameters. The continual learning stage applies Forget-Aware Replay (FAR) to estimate per-sample forgetting risk and prioritize vulnerable knowledge for rehearsal. Finally, the deployment stage employs Anchor-Protected Progressive Merging (APPM) to consolidate adapters into a single model. Experiments on DIVE confirm the framework's effectiveness in evolving blockchain ecosystems. --- Why it matters: This work matters because it tackles key challenges in smart contract vulnerability detection using LLMs, which is crucial for ensuring the security of blockchain-based systems. The proposed pipeline has the potential to improve the efficiency and accuracy of vulnerability detection in real-world applications. Source: https://arxiv.org/abs/2608.19680