Announcing Native BM25 Ranking in AlloyDB and Cloud SQL
cloud.google.com·1d ago
AI Research
arXiv CS.AI·Jun 11·5 min read
ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
TL;DR
Multimodal large language models often struggle with unwanted biases and inefficiencies in learning. ASRU, which stands for Activation Steering and Reinforcement Unlearning, was developed to address these issues by steering model activations and unlearning biased information.
✦ Why It Matters
Engineers can implement ASRU to enhance model fairness and performance in their AI applications.
Key Takeaways
- Adopt ASRU for improved unlearning in MLLMs.
- Focus on generation quality alongside unlearning effectiveness.
- Utilize minimal supervision data for optimal results.
- Experiment with activation redirection techniques.
How It Works
ASRU combines activation steering and reinforcement unlearning to manage how models forget sensitive information. It first redirects model activations to induce a refusal behavior, which prevents the model from recalling certain data.
Then, it fine-tunes the boundaries of this refusal using a customized reward function, balancing the need to forget with the necessity of maintaining high-quality outputs.