Fuzzing Large Language Models to Elicit Hidden Behaviours

1. 1Inject Gaussian noise into a model's weights using Python libraries like NumPy
2. 2Perturb residual-stream activations to analyze the model's behavior
3. 3Compare the outputs of the perturbed models to identify hidden behaviors
4. 4Run multiple iterations of fuzzing to increase the chances of eliciting the hidden behavior
5. 5Analyze the results using statistical methods to determine the significance of the findings

AI engineers and researchers can benefit from this technique to test and validate the security of their language models, while data scientists can apply it to identify potential vulnerabilities

💡 Fuzzing can be an effective technique for eliciting hidden behaviors in large language models, even without knowing the trigger

🚨 Fuzzing large language models can reveal hidden behaviors! 🤖