Interpretable sparse fine-tuning with TopKLoRA enables faithful circuit discovery
Finding and removing hidden sabotage planted in AI models through fine-tuning
Researchers developed TopKLoRA, a method to identify exactly which parts of an AI model implement hidden backdoors inserted during fine-tuning on poisoned data. They successfully located and removed these sabotage circuits from three different models while keeping the models' normal abilities intact — with backdoor circuits being up to 10 times smaller and more interpretable than previous methods.
AI systems trained on data from untrusted sources can be secretly weaponized through backdoors that persist even after safety retraining. This method lets developers actually find and surgically remove the malicious code rather than hoping it's gone, making it possible to safely use third-party datasets and validate that trained models don't contain hidden threats.