Invited Talk: Decoding-time experts for language model adaptation (Alisa Liu)
Alisa Liu
2024 Invited Talk
in
Competition: LLM Merging: Building LLMs Efficiently through Merging
in
Competition: LLM Merging: Building LLMs Efficiently through Merging
Abstract
What can we achieve by mixing the logits of multiple models at inference time? In this talk I will briefly discuss decoding-time experts (2021), where we introduced expert and anti-expert LMs for inference-time guidance, then go into depth on our recent work proxy-tuning (2024), where we apply this idea to achieve inference-time tuning of black box language models. Along the way, I will also briefly survey other works on this topic.
Video
Chat is not available.
Successful Page Load