CASD: Enhancing Generation Accuracy via Context-Aware Speculative Decoding
Keywords: Language model, Natural language processing, Long context generation, Speculative decoding
TL;DR: This paper introduces a general decoding algorithm to enhance the generation accuracy and efficiency of LMs.
Abstract: With recent advancements in long-context model variants, Large language models (LLMs) can conveniently process different types of task-related information by simply converting them into an input sequence, even consisting of over 100K tokens. Though with a simple and unified form, there is still considerable room in leveraging input context effectively and efficiently. In this paper, we propose a simple yet effective CASD (Context-Aware Speculative Decoding) method to boost context usage. CASD is a decoding algorithm that requires no extra training or draft models. It improves not only generation performance but also inference efficiency. Experiments on 8 datasets (including question answering, summarization and code completion tasks in LongBench) show that CASD increases the average generation score by 3.3 points. CASD achieves a mean acceptance length of 3.10 and a speed-up ratio of 1.99. Moreover, CASD integrates effectively with context compression technology, addressing the issue of excessive memory overhead caused by long contexts. Since CASD directly retrieves token-level content from the input context to boost the generation accuracy, it can effectively mitigate the possible side-effects of context compression methods when crucial context information is dropped. Our anonymous code is available at \href{https://anonymous.4open.science/r/CASD}{https://anonymous.4open.science/r/CASD}.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13367
Loading