Summary of Uncertainty-aware Hybrid Inference with On-device Small and Remote Large Language Models, by Seungeun Oh et al.
Uncertainty-Aware Hybrid Inference with On-Device Small and Remote Large Language Models
by Seungeun Oh, Jinhyuk Kim, Jihong Park, Seung-Woo Ko, Tony Q. S. Quek, Seong-Lyun Kim
First submitted to arxiv on: 17 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Distributed, Parallel, and Cluster Computing (cs.DC); Information Theory (cs.IT); Networking and Internet Architecture (cs.NI); Signal Processing (eess.SP)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A hybrid language model (HLM) architecture is proposed, which integrates a small language model (SLM) operating on a mobile device with a large language model (LLM) hosted at the base station of a wireless network. The SLM’s vocabulary distribution is uploaded to the LLM, which accepts or rejects it based on the speculative inference principle. However, this approach suffers from low token throughput due to uplink transmission and computation costs. To address this, an Uncertainty-aware opportunistic HLM (U-HLM) structure is proposed, where the SLM locally measures its output uncertainty and skips both uplink transmissions and LLM operations for tokens likely to be accepted. This is enabled by a linear correlation between the SLM’s uncertainty and the LLM’s rejection probability. The paper analytically derives the uncertainty threshold and evaluates its expected risk of rejection. Simulations show that U-HLM reduces uplink transmissions and LLM computations by 45.93%, while achieving up to 97.54% of the LLM’s inference accuracy and 2.54 times faster token throughput than HLM without skipping. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper explores a new way to combine small and large language models to make more accurate predictions. The idea is to use a smaller model on a mobile device to generate words, which are then sent to a larger model in the cloud for processing. However, this approach has some limitations, like taking up too much data and computation power. To fix this, researchers created a new architecture that allows the small model to skip sending certain words if it’s likely they’ll be rejected by the large model. This saves time and data and still gives good results. |
Keywords
» Artificial intelligence » Inference » Language model » Large language model » Probability » Token
