Skip to yearly menu bar Skip to main content


Invited Talk
in
Competition: LLM Merging: Building LLMs Efficiently through Merging

Invited Talk: Decoding-time experts for language model adaptation (Alisa Liu)

Alisa Liu

[ ]
Sun 15 Dec 11 a.m. PST — 11:45 a.m. PST

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.

Live content is unavailable. Log in and register to view live content