Infusion: Shaping Model Behavior by Editing Training Data via Influence Functions
📰 ArXiv cs.AI
Infusion framework uses influence functions to edit training data and induce targeted changes in model behavior
Action Steps
- Identify the desired model behavior change
- Compute influence functions to attribute model behavior to training documents
- Apply scalable influence-function approximations to find small perturbations to training documents
- Evaluate the effectiveness of Infusion on data poisoning tasks across different domains
Who Needs to Know This
ML researchers and engineers on a team can benefit from Infusion to refine model behavior, while data scientists can apply it to improve model performance
Key Insight
💡 Influence functions can be used in reverse to craft training data that induces targeted model behavior
Share This
🚀 Infusion: edit training data to shape model behavior with influence functions!
Key Takeaways
Infusion framework uses influence functions to edit training data and induce targeted changes in model behavior
Full Article
Title: Infusion: Shaping Model Behavior by Editing Training Data via Influence Functions
Abstract:
arXiv:2602.09987v4 Announce Type: replace-cross Abstract: Influence functions are commonly used to attribute model behavior to training documents. We explore the reverse: crafting training data that induces model behavior. Our framework, Infusion, uses scalable influence-function approximations to compute small perturbations to training documents that induce targeted changes in model behavior through parameter shifts. We evaluate Infusion on data poisoning tasks across vision and language domain
Abstract:
arXiv:2602.09987v4 Announce Type: replace-cross Abstract: Influence functions are commonly used to attribute model behavior to training documents. We explore the reverse: crafting training data that induces model behavior. Our framework, Infusion, uses scalable influence-function approximations to compute small perturbations to training documents that induce targeted changes in model behavior through parameter shifts. We evaluate Infusion on data poisoning tasks across vision and language domain
DeepCamp AI