We describe KVLink, an approach for efficient key-value (KV) cache reuse in large language models (LLMs). In many LLM applications, different inputs can share overlapping context, such as the same retrieved document appearing in multiple queries. However, the LLMs still need to encode the entire context for each query, leading to redundant computation. In this paper, we propose a new strategy to eliminate such inefficiency, where the KV cache of each document is precomputed independently. During inference, the KV caches of retrieved documents are concatenated, allowing the model to reuse cached representations instead of recomputing them. To mitigate the performance degradation of LLMs when using KV caches computed independently for each document, KVLink introduces three key components: adjusting positional embeddings of the KV cache at inference to match the global position after concatenation, using trainable special tokens to restore self-attention across independently encoded documents, and applying mixed-data fine-tuning to enhance performance while preserving the model's original capabilities. Experiments across 7 datasets demonstrate that KVLink improves question answering accuracy by an average of 4% over state-of-the-art methods. Furthermore, by leveraging precomputed KV caches, our approach reduces time-to-first-token by up to 90% compared to standard LLM inference, making it a scalable and efficient solution for context reuse.

Self-supervised approaches such as Momentum Contrast (MoCo) can leverage unlabeled data to produce pretrained models for subsequent fine-tuning on labeled data. While MoCo has demonstrated promising results on natural image classification tasks, its application to medical imaging tasks like chest X-ray interpretation has been limited. Chest X-ray interpretation is fundamentally different from natural image classification in ways that may limit the applicability of self-supervised approaches. In this work, we investigate whether MoCo-pretraining leads to better representations or initializations for chest X-ray interpretation. We conduct MoCo-pretraining on CheXpert, a large labeled dataset of X-rays, followed by supervised fine-tuning experiments on the pleural effusion task. Using 0.1% of labeled training data, we find that a linear model trained on MoCo-pretrained representations outperforms one trained on representations without MoCo-pretraining by an AUC of 0.096 (95% CI 0.061, 0.130), indicating that MoCo-pretrained representations are of higher quality. Furthermore, a model fine-tuned end-to-end with MoCo-pretraining outperforms its non-MoCo-pretrained counterpart by an AUC of 0.037 (95% CI 0.015, 0.062) with the 0.1% label fraction. These AUC improvements are observed for all label fractions for both the linear model and an end-to-end fine-tuned model with the greater improvements for smaller label fractions. Finally, we observe similar results on a small, target chest X-ray dataset (Shenzhen dataset for tuberculosis) with MoCo-pretraining done on the source dataset (CheXpert), which suggests that pretraining on unlabeled X-rays can provide transfer learning benefits for a target task. Our study demonstrates that MoCo-pretraining provides high-quality representations and transferable initializations for chest X-ray interpretation.