Boosting In-Context Learning in LLMs with Retrieval-based Codebook
Keywords: large language model, in-context learning, retrieval, codebook
TL;DR: This paper presents a novel method to enhance in-context learning (ICL) in large language models (LLMs) using transformer key-value pairs and a retrieval-based codebook for improved performance.
Abstract: Recent advancements in large language models (LLMs) have demonstrated exceptional performance across various downstream tasks, particularly due to their in-context learning (ICL) abilities. ICL enables models to learn from a few demonstrations presented in the context, without requiring retraining or fine-tuning. However, the effectiveness of ICL is highly dependent on factors such as prompt design and input length. To address these limitations, we propose a novel approach that leverages the key-value pairs within Transformers to enhance contextual understanding in LLMs. Specifically, our method converts raw demonstrations into task vectors—comprising keys and values—which are derived through multiple passes of the LLM, then integrated with test task vectors to improve model comprehension of the input. Furthermore, we introduce a retrieval-based codebook mechanism that captures information from long-context demonstrations while filtering irrelevant content. This codebook dynamically stores and updates task vectors generated during inference, mitigating input length constraints and optimizing the relevance of contextual data. By retrieving the most pertinent historical task vectors, the codebook ensures that only relevant information is utilized during inference. Extensive experiments show that these enhancements significantly outperform conventional ICL, achieving superior accuracy and efficiency. Overall, this work sets a new benchmark for optimizing ICL in LLMs, enabling their effective deployment in complex, real-world applications.
Primary Area: foundation or frontier models, including LLMs
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.
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: 10206
Loading