This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
CoSteer: Decoding-Time Personalization through Localized Delta Steering
Outline
This paper proposes CoSteer, a novel framework for personalized text generation. CoSteer adjusts the output of a cloud-based LLM by leveraging the difference in logits between personalized and non-personalized text generated by a local small model, enabling real-time adaptation under resource constraints on personal devices. This method dynamically adjusts the delta vector locally, preserving privacy while maintaining the general capabilities of cloud LLM.
Takeaways, Limitations
•
Takeaways:
◦
Leveraging user profiles and records on personal devices effectively supports LLM in generating personalized content.
◦
On-device data processing ensures privacy.
◦
Keeps computational overhead to an acceptable level while maintaining the general capabilities of cloud-based LLM.
◦
Enhances privacy by transmitting only the final output without transferring any data.
•
Limitations:
◦
The paper does not mention anything specific about Limitations.
