More Tokens, Lower Precision: Towards the Optimal Token-Precision Trade-off in KV Cache Compression

As large language models (LLMs) process increasing context windows, the memory usage of KV cache has become a critical bottleneck during inference. The mainstream KV compression methods, including KV pruning and KV quantization, primarily focus on either token or precision dimension and seldom explore the efficiency of their combination. In this paper, we comprehensively investigate the token-precision trade-off in KV cache compression. Experiments demonstrate that storing more tokens in the KV cache with lower precision, i.e., quantized pruning, can significantly enhance the long-context performance of LLMs. Furthermore, in-depth analysis regarding token-precision trade-off from a series of key aspects exhibit that, quantized pruning achieves substantial improvements in retrieval-related tasks and consistently performs well across varying input lengths. Moreover, quantized pruning demonstrates notable stability across different KV pruning methods, quantization strategies, and model scales. These findings provide valuable insights into the token-precision trade-off in KV cache compression. We plan to release our code in the near future.

Quasi-Newton OMP Approach for Super-Resolution Channel Estimation and Extrapolation

Channel estimation and extrapolation are fundamental issues in MIMO communication systems. In this paper, we proposed the quasi-Newton orthogonal matching pursuit (QNOMP) approach to overcome these issues with high efficiency while maintaining accuracy. The algorithm consists of two stages on the super-resolution recovery: we first performed a cheap on-grid OMP estimation of channel parameters in the sparsity domain (e.g., delay or angle), then an off-grid optimization to achieve the super-resolution. In the off-grid stage, we employed the BFGS quasi-Newton method to jointly estimate the parameters through a multipath model, which improved the speed and accuracy significantly. Furthermore, we derived the optimal extrapolated solution in the linear minimum mean squared estimator criterion, revealed its connection with Slepian basis, and presented a practical algorithm to realize the extrapolation based on the QNOMP results. Special treatment utilizing the block sparsity nature of the considered channels was also proposed. Numerical experiments on the simulated models and CDL-C channels demonstrated the high performance and low computational complexity of QNOMP.

DMRA: An Adaptive Line Spectrum Estimation Method through Dynamical Multi-Resolution of Atoms

We proposed a novel dense line spectrum super-resolution algorithm, the DMRA, that leverages dynamical multi-resolution of atoms technique to address the limitation of traditional compressed sensing methods when handling dense point-source signals. The algorithm utilizes a smooth $\tanh$ relaxation function to replace the $\ell_0$ norm, promoting sparsity and jointly estimating the frequency atoms and complex gains. To reduce computational complexity and improve frequency estimation accuracy, a two-stage strategy was further introduced to dynamically adjust the number of the optimized degrees of freedom. The strategy first increases candidate frequencies through local refinement, then applies a sparse selector to eliminate insignificant frequencies, thereby adaptively adjusting the degrees of freedom to improve estimation accuracy. Theoretical analysis were provided to validate the proposed method for multi-parameter estimations. Computational results demonstrated that this algorithm achieves good super-resolution performance in various practical scenarios and outperforms the state-of-the-art methods in terms of frequency estimation accuracy and computational efficiency.

DePrompt: Desensitization and Evaluation of Personal Identifiable Information in Large Language Model Prompts

Prompt serves as a crucial link in interacting with large language models (LLMs), widely impacting the accuracy and interpretability of model outputs. However, acquiring accurate and high-quality responses necessitates precise prompts, which inevitably pose significant risks of personal identifiable information (PII) leakage. Therefore, this paper proposes DePrompt, a desensitization protection and effectiveness evaluation framework for prompt, enabling users to safely and transparently utilize LLMs. Specifically, by leveraging large model fine-tuning techniques as the underlying privacy protection method, we integrate contextual attributes to define privacy types, achieving high-precision PII entity identification. Additionally, through the analysis of key features in prompt desensitization scenarios, we devise adversarial generative desensitization methods that retain important semantic content while disrupting the link between identifiers and privacy attributes. Furthermore, we present utility evaluation metrics for prompt to better gauge and balance privacy and usability. Our framework is adaptable to prompts and can be extended to text usability-dependent scenarios. Through comparison with benchmarks and other model methods, experimental evaluations demonstrate that our desensitized prompt exhibit superior privacy protection utility and model inference results.

Leveraging Adaptive Implicit Representation Mapping for Ultra High-Resolution Image Segmentation

Implicit representation mapping (IRM) can translate image features to any continuous resolution, showcasing its potent capability for ultra-high-resolution image segmentation refinement. Current IRM-based methods for refining ultra-high-resolution image segmentation often rely on CNN-based encoders to extract image features and apply a Shared Implicit Representation Mapping Function (SIRMF) to convert pixel-wise features into segmented results. Hence, these methods exhibit two crucial limitations. Firstly, the CNN-based encoder may not effectively capture long-distance information, resulting in a lack of global semantic information in the pixel-wise features. Secondly, SIRMF is shared across all samples, which limits its ability to generalize and handle diverse inputs. To address these limitations, we propose a novel approach that leverages the newly proposed Adaptive Implicit Representation Mapping (AIRM) for ultra-high-resolution Image Segmentation. Specifically, the proposed method comprises two components: (1) the Affinity Empowered Encoder (AEE), a robust feature extractor that leverages the benefits of the transformer architecture and semantic affinity to model long-distance features effectively, and (2) the Adaptive Implicit Representation Mapping Function (AIRMF), which adaptively translates pixel-wise features without neglecting the global semantic information, allowing for flexible and precise feature translation. We evaluated our method on the commonly used ultra-high-resolution segmentation refinement datasets, i.e., BIG and PASCAL VOC 2012. The extensive experiments demonstrate that our method outperforms competitors by a large margin. The code is provided in supplementary material.

CLII: Visual-Text Inpainting via Cross-Modal Predictive Interaction

Image inpainting aims to fill missing pixels in damaged images and has achieved significant progress with cut-edging learning techniques. Nevertheless, state-of-the-art inpainting methods are mainly designed for nature images and cannot correctly recover text within scene text images, and training existing models on the scene text images cannot fix the issues. In this work, we identify the visual-text inpainting task to achieve high-quality scene text image restoration and text completion: Given a scene text image with unknown missing regions and the corresponding text with unknown missing characters, we aim to complete the missing information in both images and text by leveraging their complementary information. Intuitively, the input text, even if damaged, contains language priors of the contents within the images and can guide the image inpainting. Meanwhile, the scene text image includes the appearance cues of the characters that could benefit text recovery. To this end, we design the cross-modal predictive interaction (CLII) model containing two branches, i.e., ImgBranch and TxtBranch, for scene text inpainting and text completion, respectively while leveraging their complementary effectively. Moreover, we propose to embed our model into the SOTA scene text spotting method and significantly enhance its robustness against missing pixels, which demonstrates the practicality of the newly developed task. To validate the effectiveness of our method, we construct three real datasets based on existing text-related datasets, containing 1838 images and covering three scenarios with curved, incidental, and styled texts, and conduct extensive experiments to show that our method outperforms baselines significantly.

EWEK-QA: Enhanced Web and Efficient Knowledge Graph Retrieval for Citation-based Question Answering Systems

The emerging citation-based QA systems are gaining more attention especially in generative AI search applications. The importance of extracted knowledge provided to these systems is vital from both accuracy (completeness of information) and efficiency (extracting the information in a timely manner). In this regard, citation-based QA systems are suffering from two shortcomings. First, they usually rely only on web as a source of extracted knowledge and adding other external knowledge sources can hamper the efficiency of the system. Second, web-retrieved contents are usually obtained by some simple heuristics such as fixed length or breakpoints which might lead to splitting information into pieces. To mitigate these issues, we propose our enhanced web and efficient knowledge graph (KG) retrieval solution (EWEK-QA) to enrich the content of the extracted knowledge fed to the system. This has been done through designing an adaptive web retriever and incorporating KGs triples in an efficient manner. We demonstrate the effectiveness of EWEK-QA over the open-source state-of-the-art (SoTA) web-based and KG baseline models using a comprehensive set of quantitative and human evaluation experiments. Our model is able to: first, improve the web-retriever baseline in terms of extracting more relevant passages (>20\%), the coverage of answer span (>25\%) and self containment (>35\%); second, obtain and integrate KG triples into its pipeline very efficiently (by avoiding any LLM calls) to outperform the web-only and KG-only SoTA baselines significantly in 7 quantitative QA tasks and our human evaluation.

Evaluating the External and Parametric Knowledge Fusion of Large Language Models

Integrating external knowledge into large language models (LLMs) presents a promising solution to overcome the limitations imposed by their antiquated and static parametric memory. Prior studies, however, have tended to over-reliance on external knowledge, underestimating the valuable contributions of an LLMs' intrinsic parametric knowledge. The efficacy of LLMs in blending external and parametric knowledge remains largely unexplored, especially in cases where external knowledge is incomplete and necessitates supplementation by their parametric knowledge. We propose to deconstruct knowledge fusion into four distinct scenarios, offering the first thorough investigation of LLM behavior across each. We develop a systematic pipeline for data construction and knowledge infusion to simulate these fusion scenarios, facilitating a series of controlled experiments. Our investigation reveals that enhancing parametric knowledge within LLMs can significantly bolster their capability for knowledge integration. Nonetheless, we identify persistent challenges in memorizing and eliciting parametric knowledge, and determining parametric knowledge boundaries. Our findings aim to steer future explorations on harmonizing external and parametric knowledge within LLMs.

Retrieval-based Full-length Wikipedia Generation for Emergent Events

In today's fast-paced world, the growing demand to quickly generate comprehensive and accurate Wikipedia documents for emerging events is both crucial and challenging. However, previous efforts in Wikipedia generation have often fallen short of meeting real-world requirements. Some approaches focus solely on generating segments of a complete Wikipedia document, while others overlook the importance of faithfulness in generation or fail to consider the influence of the pre-training corpus. In this paper, we simulate a real-world scenario where structured full-length Wikipedia documents are generated for emergent events using input retrieved from web sources. To ensure that Large Language Models (LLMs) are not trained on corpora related to recently occurred events, we select events that have taken place recently and introduce a new benchmark Wiki-GenBen, which consists of 309 events paired with their corresponding retrieved web pages for generating evidence. Additionally, we design a comprehensive set of systematic evaluation metrics and baseline methods, to evaluate the capability of LLMs in generating factual full-length Wikipedia documents. The data and code are open-sourced at WikiGenBench.

Does the Generator Mind its Contexts? An Analysis of Generative Model Faithfulness under Context Transfer

The present study introduces the knowledge-augmented generator, which is specifically designed to produce information that remains grounded in contextual knowledge, regardless of alterations in the context. Previous research has predominantly focused on examining hallucinations stemming from static input, such as in the domains of summarization or machine translation. However, our investigation delves into the faithfulness of generative question answering in the presence of dynamic knowledge. Our objective is to explore the existence of hallucinations arising from parametric memory when contextual knowledge undergoes changes, while also analyzing the underlying causes for their occurrence. In order to efficiently address this issue, we propose a straightforward yet effective measure for detecting such hallucinations. Intriguingly, our investigation uncovers that all models exhibit a tendency to generate previous answers as hallucinations. To gain deeper insights into the underlying causes of this phenomenon, we conduct a series of experiments that verify the critical role played by context in hallucination, both during training and testing, from various perspectives.