Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 28·5 min read
What the "Spotless" Mind Remembers: How Knowledge Entanglement Shapes What Leaks After Unlearning in LLMs
TL;DR
Research reveals that the structural entanglement of knowledge in large language models (LLMs) significantly influences what information leaks after unlearning. Notably, different unlearning algorithms affect this relationship, with one algorithm reversing the expected outcomes.
✦ Why It Matters
Evaluate your LLM's knowledge structure to identify prompts that may leak sensitive information after unlearning.
Key Takeaways
- Assess knowledge entanglement to predict unlearning outcomes.
- Use WHP and GA+KL to explore unlearning effects.
- Implement the predictive model for auditing LLM prompts.
- Investigate the causal relationships in knowledge recall.
How It Works
The study manipulates a prompt's entanglement score while keeping its content constant, allowing researchers to observe how changes in entanglement affect recall. This manipulation provides evidence that unlearning impacts the underlying knowledge structure rather than just the output.