Xuefei Ning

#ICML2024 Spotlight <Differentially Private Synthetic Data via Foundation Model APIs 2: Text> was accepted at ICML 2024 as Spotlight! Unfortunately, we are not attending ICML in person. But feel free to reach out to us if you are interested! Paper: arxiv.org/abs/2403.01749

Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding paper page: huggingface.co/papers/2307.15… This work aims at decreasing the end-to-end generation latency of large language models (LLMs). One of the major causes of the high generation latency is the sequential decoding approach adopted by almost all state-of-the-art LLMs. In this work, motivated by the thinking and writing process of humans, we propose "Skeleton-of-Thought" (SoT), which guides LLMs to first generate the skeleton of the answer, and then conducts parallel API calls or batched decoding to complete the contents of each skeleton point in parallel. Not only does SoT provide considerable speed-up (up to 2.39x across 11 different LLMs), but it can also potentially improve the answer quality on several question categories in terms of diversity and relevance. SoT is an initial attempt at data-centric optimization for efficiency, and reveal the potential of pushing LLMs to think more like a human for answer quality.

Skeleton-of-Thought: LLMs can do parallel decoding Interesting prompting strategy which firsts generate an answer skeleton and then performs parallel API calls to generate the content of each skeleton point. Reports quality improvements in addition to speed-up of up to 2.39x. Big deal given how costly in terms of latency some tasks are. This a great paper to rethink the necessity of sequential decoding of current LLMs. arxiv.org/abs/2307.15337

Slides for my latest NAS tutorial with @crwhite_ml at #AutoML_Conf are now online: Part 1: crwhite.ml/assets/nas_tut… Part 2: crwhite.ml/assets/nas_tut…. Video links coming soon.