Upload data/train.jsonl with huggingface_hub

data/train.jsonl

@@ -0,0 +1,63 @@
+{"qid": "2506.02838v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nEconomic inequality is a global challenge, intensifying disparities in\neducation, healthcare, and social stability. Traditional systems like the U.S.\nfederal income tax reduce inequality but lack adaptability. Although models\nlike the Saez Optimal Taxation adjust dynamically, they fail to address\ntaxpayer heterogeneity and irrational behavior. This study introduces TaxAgent,\na novel integration of large language models (LLMs) with agent-based modeling\n(ABM) to design adaptive tax policies. In our macroeconomic simulation,\nheterogeneous H-Agents (households) simulate real-world taxpayer behaviors\nwhile the TaxAgent (government) utilizes LLMs to iteratively optimize tax\nrates, balancing equity and productivity. Benchmarked against Saez Optimal\nTaxation, U.S. federal income taxes, and free markets, TaxAgent achieves\nsuperior equity-efficiency trade-offs. This research offers a novel taxation\nsolution and a scalable, data-driven framework for fiscal policy evaluation."}
+{"qid": "2506.02634v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nServing large language models (LLMs) is important for cloud providers, and\ncaching intermediate results (KV\\$) after processing each request substantially\nimproves serving throughput and latency. However, there is limited\nunderstanding of how LLM serving benefits from KV\\$ caching, where system\ndesign decisions like cache eviction policies are highly workload-dependent. In\nthis paper, we present the first systematic characterization of the KV\\$\nworkload patterns from one of the leading LLM service providers. We draw\nobservations that were not covered by previous studies focusing on synthetic\nworkloads, including: KV\\$ reuses are skewed across requests, where reuses\nbetween single-turn requests are equally important as multi-turn requests; the\nreuse time and probability are diverse considering all requests, but for a\nspecific request category, the pattern tends to be predictable; and the overall\ncache size required for an ideal cache hit ratio is moderate. Based on the\ncharacterization, we further propose a workload-aware cache eviction policy\nthat improves the serving performance under real-world traces, especially with\nlimited cache capacity."}
+{"qid": "2506.00958v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nNonverbal communication is integral to human interaction, with gestures,\nfacial expressions, and body language conveying critical aspects of intent and\nemotion. However, existing large language models (LLMs) fail to effectively\nincorporate these nonverbal elements, limiting their capacity to create fully\nimmersive conversational experiences. We introduce MARS, a multimodal language\nmodel designed to understand and generate nonverbal cues alongside text,\nbridging this gap in conversational AI. Our key innovation is VENUS, a\nlarge-scale dataset comprising annotated videos with time-aligned text, facial\nexpressions, and body language. Leveraging VENUS, we train MARS with a\nnext-token prediction objective, combining text with vector-quantized nonverbal\nrepresentations to achieve multimodal understanding and generation within a\nunified framework. Based on various analyses of the VENUS datasets, we validate\nits substantial scale and high effectiveness. Our quantitative and qualitative\nresults demonstrate that MARS successfully generates text and nonverbal\nlanguages, corresponding to conversational input."}
+{"qid": "2506.00832v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nRecent advances in Text-to-Speech (TTS) have significantly improved speech\nnaturalness, increasing the demand for precise prosody control and\nmispronunciation correction. Existing approaches for prosody manipulation often\ndepend on specialized modules or additional training, limiting their capacity\nfor post-hoc adjustments. Similarly, traditional mispronunciation correction\nrelies on grapheme-to-phoneme dictionaries, making it less practical in\nlow-resource settings. We introduce Counterfactual Activation Editing, a\nmodel-agnostic method that manipulates internal representations in a\npre-trained TTS model to achieve post-hoc control of prosody and pronunciation.\nExperimental results show that our method effectively adjusts prosodic features\nand corrects mispronunciations while preserving synthesis quality. This opens\nthe door to inference-time refinement of TTS outputs without retraining,\nbridging the gap between pre-trained TTS models and editable speech synthesis."}
+{"qid": "2506.00418v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn context learning (ICL) relies heavily on high quality demonstrations drawn\nfrom large annotated corpora. Existing approaches detect noisy annotations by\nranking local perplexities, presuming that noisy samples yield higher\nperplexities than their clean counterparts. However, this assumption breaks\ndown when the noise ratio is high and many demonstrations are flawed. We\nreexamine the perplexity based paradigm for text generation under noisy\nannotations, highlighting two sources of bias in perplexity: the annotation\nitself and the domain specific knowledge inherent in large language models\n(LLMs). To overcome these biases, we introduce a dual debiasing framework that\nuses synthesized neighbors to explicitly correct perplexity estimates, yielding\na robust Sample Cleanliness Score. This metric uncovers absolute sample\ncleanliness regardless of the overall corpus noise level. Extensive experiments\ndemonstrate our method's superior noise detection capabilities and show that\nits final ICL performance is comparable to that of a fully clean demonstration\ncorpus. Moreover, our approach remains robust even when noise ratios are\nextremely high."}
+{"qid": "2505.24754v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn this work, we investigate an important task named instruction-following\ntext embedding, which generates dynamic text embeddings that adapt to user\ninstructions, highlighting specific attributes of text. Despite recent\nadvancements, existing approaches suffer from significant computational\noverhead, as they require re-encoding the entire corpus for each new\ninstruction. To address this challenge, we propose GSTransform, a novel\ninstruction-following text embedding framework based on Guided Space\nTransformation. Our key observation is that instruction-relevant information is\ninherently encoded in generic embeddings but remains underutilized. Instead of\nrepeatedly encoding the corpus for each instruction, GSTransform is a\nlightweight transformation mechanism that adapts pre-computed embeddings in\nreal time to align with user instructions, guided by a small amount of text\ndata with instruction-focused label annotation. We conduct extensive\nexperiments on three instruction-awareness downstream tasks across nine\nreal-world datasets, demonstrating that GSTransform improves\ninstruction-following text embedding quality over state-of-the-art methods\nwhile achieving dramatic speedups of 6~300x in real-time processing on\nlarge-scale datasets. The source code is available at\nhttps://github.com/YingchaojieFeng/GSTransform."}
+{"qid": "2505.24575v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nSummarizing long-form narratives--such as books, movies, and TV\nscripts--requires capturing intricate plotlines, character interactions, and\nthematic coherence, a task that remains challenging for existing LLMs. We\nintroduce NexusSum, a multi-agent LLM framework for narrative summarization\nthat processes long-form text through a structured, sequential\npipeline--without requiring fine-tuning. Our approach introduces two key\ninnovations: (1) Dialogue-to-Description Transformation: A narrative-specific\npreprocessing method that standardizes character dialogue and descriptive text\ninto a unified format, improving coherence. (2) Hierarchical Multi-LLM\nSummarization: A structured summarization pipeline that optimizes chunk\nprocessing and controls output length for accurate, high-quality summaries. Our\nmethod establishes a new state-of-the-art in narrative summarization, achieving\nup to a 30.0% improvement in BERTScore (F1) across books, movies, and TV\nscripts. These results demonstrate the effectiveness of multi-agent LLMs in\nhandling long-form content, offering a scalable approach for structured\nsummarization in diverse storytelling domains."}
+{"qid": "2506.00085v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nLarge Language Models (LLMs) encode behaviors such as refusal within their\nactivation space, yet identifying these behaviors remains a significant\nchallenge. Existing methods often rely on predefined refusal templates\ndetectable in output tokens or require manual analysis. We introduce\n\\textbf{COSMIC} (Cosine Similarity Metrics for Inversion of Concepts), an\nautomated framework for direction selection that identifies viable steering\ndirections and target layers using cosine similarity - entirely independent of\nmodel outputs. COSMIC achieves steering performance comparable to prior methods\nwithout requiring assumptions about a model's refusal behavior, such as the\npresence of specific refusal tokens. It reliably identifies refusal directions\nin adversarial settings and weakly aligned models, and is capable of steering\nsuch models toward safer behavior with minimal increase in false refusals,\ndemonstrating robustness across a wide range of alignment conditions."}
+{"qid": "2505.23996v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nThe recent rapid adoption of large language models (LLMs) highlights the\ncritical need for benchmarking their fairness. Conventional fairness metrics,\nwhich focus on discrete accuracy-based evaluations (i.e., prediction\ncorrectness), fail to capture the implicit impact of model uncertainty (e.g.,\nhigher model confidence about one group over another despite similar accuracy).\nTo address this limitation, we propose an uncertainty-aware fairness metric,\nUCerF, to enable a fine-grained evaluation of model fairness that is more\nreflective of the internal bias in model decisions compared to conventional\nfairness measures. Furthermore, observing data size, diversity, and clarity\nissues in current datasets, we introduce a new gender-occupation fairness\nevaluation dataset with 31,756 samples for co-reference resolution, offering a\nmore diverse and suitable dataset for evaluating modern LLMs. We establish a\nbenchmark, using our metric and dataset, and apply it to evaluate the behavior\nof ten open-source LLMs. For example, Mistral-7B exhibits suboptimal fairness\ndue to high confidence in incorrect predictions, a detail overlooked by\nEqualized Odds but captured by UCerF. Overall, our proposed LLM benchmark,\nwhich evaluates fairness with uncertainty awareness, paves the way for\ndeveloping more transparent and accountable AI systems."}
+{"qid": "2505.23353v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nQuantitative susceptibility maps from magnetic resonance images can provide\nboth prognostic and diagnostic information in multiple sclerosis, a\nneurodegenerative disease characterized by the formation of lesions in white\nmatter brain tissue. In particular, susceptibility maps provide adequate\ncontrast to distinguish between \"rim\" lesions, surrounded by deposited\nparamagnetic iron, and \"non-rim\" lesion types. These paramagnetic rim lesions\n(PRLs) are an emerging biomarker in multiple sclerosis. Much effort has been\ndevoted to both detection and segmentation of such lesions to monitor\nlongitudinal change. As paramagnetic rim lesions are rare, addressing this\nproblem requires confronting the class imbalance between rim and non-rim\nlesions. We produce synthetic quantitative susceptibility maps of paramagnetic\nrim lesions and show that inclusion of such synthetic data improves classifier\nperformance and provide a multi-channel extension to generate accompanying\ncontrasts and probabilistic segmentation maps. We exploit the projection\ncapability of our trained generative network to demonstrate a novel denoising\napproach that allows us to train on ambiguous rim cases and substantially\nincrease the minority class. We show that both synthetic lesion synthesis and\nour proposed rim lesion label denoising method best approximate the unseen rim\nlesion distribution and improve detection in a clinically interpretable manner.\nWe release our code and generated data at https://github.com/agr78/PRLx-GAN\nupon publication."}
+{"qid": "2505.22757v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nMulti-token prediction (MTP) is a recently proposed pre-training objective\nfor language models. Rather than predicting only the next token (NTP), MTP\npredicts the next $k$ tokens at each prediction step, using multiple prediction\nheads. MTP has shown promise in improving downstream performance, inference\nspeed, and training efficiency, particularly for large models. However, prior\nwork has shown that smaller language models (SLMs) struggle with the MTP\nobjective. To address this, we propose a curriculum learning strategy for MTP\ntraining, exploring two variants: a forward curriculum, which gradually\nincreases the complexity of the pre-training objective from NTP to MTP, and a\nreverse curriculum, which does the opposite. Our experiments show that the\nforward curriculum enables SLMs to better leverage the MTP objective during\npre-training, improving downstream NTP performance and generative output\nquality, while retaining the benefits of self-speculative decoding. The reverse\ncurriculum achieves stronger NTP performance and output quality, but fails to\nprovide any self-speculative decoding benefits."}
+{"qid": "2506.02853v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nAction coordination in human structure is indispensable for the spatial\nconstraints of 2D joints to recover 3D pose. Usually, action coordination is\nrepresented as a long-range dependence among body parts. However, there are two\nmain challenges in modeling long-range dependencies. First, joints should not\nonly be constrained by other individual joints but also be modulated by the\nbody parts. Second, existing methods make networks deeper to learn dependencies\nbetween non-linked parts. They introduce uncorrelated noise and increase the\nmodel size. In this paper, we utilize a pyramid structure to better learn\npotential long-range dependencies. It can capture the correlation across joints\nand groups, which complements the context of the human sub-structure. In an\neffective cross-scale way, it captures the pyramid-structured long-range\ndependence. Specifically, we propose a novel Pyramid Graph Attention (PGA)\nmodule to capture long-range cross-scale dependencies. It concatenates\ninformation from various scales into a compact sequence, and then computes the\ncorrelation between scales in parallel. Combining PGA with graph convolution\nmodules, we develop a Pyramid Graph Transformer (PGFormer) for 3D human pose\nestimation, which is a lightweight multi-scale transformer architecture. It\nencapsulates human sub-structures into self-attention by pooling. Extensive\nexperiments show that our approach achieves lower error and smaller model size\nthan state-of-the-art methods on Human3.6M and MPI-INF-3DHP datasets. The code\nis available at https://github.com/MingjieWe/PGFormer."}
+{"qid": "2506.02547v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nEvent cameras capture scene changes asynchronously on a per-pixel basis,\nenabling extremely high temporal resolution. However, this advantage comes at\nthe cost of high bandwidth, memory, and computational demands. To address this,\nprior work has explored event downsampling, but most approaches rely on fixed\nheuristics or threshold-based strategies, limiting their adaptability. Instead,\nwe propose a probabilistic framework, POLED, that models event importance\nthrough an event-importance probability density function (ePDF), which can be\narbitrarily defined and adapted to different applications. Our approach\noperates in a purely online setting, estimating event importance on-the-fly\nfrom raw event streams, enabling scene-specific adaptation. Additionally, we\nintroduce zero-shot event downsampling, where downsampled events must remain\nusable for models trained on the original event stream, without task-specific\nadaptation. We design a contour-preserving ePDF that prioritizes structurally\nimportant events and evaluate our method across four datasets and tasks--object\nclassification, image interpolation, surface normal estimation, and object\ndetection--demonstrating that intelligent sampling is crucial for maintaining\nperformance under event-budget constraints."}
+{"qid": "2506.01071v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn this paper, we propose an Aligned Contrastive Learning (ACL) algorithm to\naddress the long-tailed recognition problem. Our findings indicate that while\nmulti-view training boosts the performance, contrastive learning does not\nconsistently enhance model generalization as the number of views increases.\nThrough theoretical gradient analysis of supervised contrastive learning (SCL),\nwe identify gradient conflicts, and imbalanced attraction and repulsion\ngradients between positive and negative pairs as the underlying issues. Our ACL\nalgorithm is designed to eliminate these problems and demonstrates strong\nperformance across multiple benchmarks. We validate the effectiveness of ACL\nthrough experiments on long-tailed CIFAR, ImageNet, Places, and iNaturalist\ndatasets. Results show that ACL achieves new state-of-the-art performance."}
+{"qid": "2506.01037v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nExisting diffusion-based video super-resolution (VSR) methods are susceptible\nto introducing complex degradations and noticeable artifacts into\nhigh-resolution videos due to their inherent randomness. In this paper, we\npropose a noise-robust real-world VSR framework by incorporating\nself-supervised learning and Mamba into pre-trained latent diffusion models. To\nensure content consistency across adjacent frames, we enhance the diffusion\nmodel with a global spatio-temporal attention mechanism using the Video\nState-Space block with a 3D Selective Scan module, which reinforces coherence\nat an affordable computational cost. To further reduce artifacts in generated\ndetails, we introduce a self-supervised ControlNet that leverages HR features\nas guidance and employs contrastive learning to extract degradation-insensitive\nfeatures from LR videos. Finally, a three-stage training strategy based on a\nmixture of HR-LR videos is proposed to stabilize VSR training. The proposed\nSelf-supervised ControlNet with Spatio-Temporal Continuous Mamba based VSR\nalgorithm achieves superior perceptual quality than state-of-the-arts on\nreal-world VSR benchmark datasets, validating the effectiveness of the proposed\nmodel design and training strategies."}
+{"qid": "2506.00434v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn this paper, we address the crucial task of brain tumor segmentation in\nmedical imaging and propose innovative approaches to enhance its performance.\nThe current state-of-the-art nnU-Net has shown promising results but suffers\nfrom extensive training requirements and underutilization of pre-trained\nweights. To overcome these limitations, we integrate Axial-Coronal-Sagittal\nconvolutions and pre-trained weights from ImageNet into the nnU-Net framework,\nresulting in reduced training epochs, reduced trainable parameters, and\nimproved efficiency. Two strategies for transferring 2D pre-trained weights to\nthe 3D domain are presented, ensuring the preservation of learned relationships\nand feature representations critical for effective information propagation.\nFurthermore, we explore a joint classification and segmentation model that\nleverages pre-trained encoders from a brain glioma grade classification proxy\ntask, leading to enhanced segmentation performance, especially for challenging\ntumor labels. Experimental results demonstrate that our proposed methods in the\nfast training settings achieve comparable or even outperform the ensemble of\ncross-validation models, a common practice in the brain tumor segmentation\nliterature."}
+{"qid": "2506.00333v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nOpen-vocabulary object detection models allow users to freely specify a class\nvocabulary in natural language at test time, guiding the detection of desired\nobjects. However, vocabularies can be overly broad or even mis-specified,\nhampering the overall performance of the detector. In this work, we propose a\nplug-and-play Vocabulary Adapter (VocAda) to refine the user-defined\nvocabulary, automatically tailoring it to categories that are relevant for a\ngiven image. VocAda does not require any training, it operates at inference\ntime in three steps: i) it uses an image captionner to describe visible\nobjects, ii) it parses nouns from those captions, and iii) it selects relevant\nclasses from the user-defined vocabulary, discarding irrelevant ones.\nExperiments on COCO and Objects365 with three state-of-the-art detectors show\nthat VocAda consistently improves performance, proving its versatility. The\ncode is open source."}
+{"qid": "2505.24443v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nConventional semi-supervised learning (SSL) ideally assumes that labeled and\nunlabeled data share an identical class distribution, however in practice, this\nassumption is easily violated, as unlabeled data often includes unknown class\ndata, i.e., outliers. The outliers are treated as noise, considerably degrading\nthe performance of SSL models. To address this drawback, we propose a novel\nframework, Diversify and Conquer (DAC), to enhance SSL robustness in the\ncontext of open-set semi-supervised learning. In particular, we note that\nexisting open-set SSL methods rely on prediction discrepancies between inliers\nand outliers from a single model trained on labeled data. This approach can be\neasily failed when the labeled data is insufficient, leading to performance\ndegradation that is worse than naive SSL that do not account for outliers. In\ncontrast, our approach exploits prediction disagreements among multiple models\nthat are differently biased towards the unlabeled distribution. By leveraging\nthe discrepancies arising from training on unlabeled data, our method enables\nrobust outlier detection even when the labeled data is underspecified. Our key\ncontribution is constructing a collection of differently biased models through\na single training process. By encouraging divergent heads to be differently\nbiased towards outliers while making consistent predictions for inliers, we\nexploit the disagreement among these heads as a measure to identify unknown\nconcepts. Our code is available at https://github.com/heejokong/DivCon."}
+{"qid": "2505.24334v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn the era of intelligent manufacturing, anomaly detection has become\nessential for maintaining quality control on modern production lines. However,\nwhile many existing models show promising performance, they are often too\nlarge, computationally demanding, and impractical to deploy on\nresource-constrained embedded devices that can be easily installed on the\nproduction lines of Small and Medium Enterprises (SMEs). To bridge this gap, we\npresent KairosAD, a novel supervised approach that uses the power of the Mobile\nSegment Anything Model (MobileSAM) for image-based anomaly detection. KairosAD\nhas been evaluated on the two well-known industrial anomaly detection datasets,\ni.e., MVTec-AD and ViSA. The results show that KairosAD requires 78% fewer\nparameters and boasts a 4x faster inference time compared to the leading\nstate-of-the-art model, while maintaining comparable AUROC performance. We\ndeployed KairosAD on two embedded devices, the NVIDIA Jetson NX, and the NVIDIA\nJetson AGX. Finally, KairosAD was successfully installed and tested on the real\nproduction line of the Industrial Computer Engineering Laboratory (ICE Lab) at\nthe University of Verona. The code is available at\nhttps://github.com/intelligolabs/KairosAD."}
+{"qid": "2505.23290v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn 3D speech-driven facial animation generation, existing methods commonly\nemploy pre-trained self-supervised audio models as encoders. However, due to\nthe prevalence of phonetically similar syllables with distinct lip shapes in\nlanguage, these near-homophone syllables tend to exhibit significant coupling\nin self-supervised audio feature spaces, leading to the averaging effect in\nsubsequent lip motion generation. To address this issue, this paper proposes a\nplug-and-play semantic decorrelation module-Wav2Sem. This module extracts\nsemantic features corresponding to the entire audio sequence, leveraging the\nadded semantic information to decorrelate audio encodings within the feature\nspace, thereby achieving more expressive audio features. Extensive experiments\nacross multiple Speech-driven models indicate that the Wav2Sem module\neffectively decouples audio features, significantly alleviating the averaging\neffect of phonetically similar syllables in lip shape generation, thereby\nenhancing the precision and naturalness of facial animations. Our source code\nis available at https://github.com/wslh852/Wav2Sem.git."}
+{"qid": "2505.23180v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDeep-unrolling and plug-and-play (PnP) approaches have become the de-facto\nstandard solvers for single-pixel imaging (SPI) inverse problem. PnP\napproaches, a class of iterative algorithms where regularization is implicitly\nperformed by an off-the-shelf deep denoiser, are flexible for varying\ncompression ratios (CRs) but are limited in reconstruction accuracy and speed.\nConversely, unrolling approaches, a class of multi-stage neural networks where\na truncated iterative optimization process is transformed into an end-to-end\ntrainable network, typically achieve better accuracy with faster inference but\nrequire fine-tuning or even retraining when CR changes. In this paper, we\naddress the challenge of integrating the strengths of both classes of solvers.\nTo this end, we design an efficient deep image restorer (DIR) for the unrolling\nof HQS (half quadratic splitting) and ADMM (alternating direction method of\nmultipliers). More importantly, a general proximal trajectory (PT) loss\nfunction is proposed to train HQS/ADMM-unrolling networks such that learned DIR\napproximates the proximal operator of an ideal explicit restoration\nregularizer. Extensive experiments demonstrate that, the resulting proximal\nunrolling networks can not only flexibly handle varying CRs with a single model\nlike PnP algorithms, but also outperform previous CR-specific unrolling\nnetworks in both reconstruction accuracy and speed. Source codes and models are\navailable at https://github.com/pwangcs/ProxUnroll."}
+{"qid": "2505.22616v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nNeural rendering methods have gained significant attention for their ability\nto reconstruct 3D scenes from 2D images. The core idea is to take multiple\nviews as input and optimize the reconstructed scene by minimizing the\nuncertainty in geometry and appearance across the views. However, the\nreconstruction quality is limited by the number of input views. This limitation\nis further pronounced in complex and dynamic scenes, where certain angles of\nobjects are never seen. In this paper, we propose to use video frame\ninterpolation as the data augmentation method for neural rendering.\nFurthermore, we design a lightweight yet high-quality video frame interpolation\nmodel, PS4PRO (Pixel-to-pixel Supervision for Photorealistic Rendering and\nOptimization). PS4PRO is trained on diverse video datasets, implicitly modeling\ncamera movement as well as real-world 3D geometry. Our model performs as an\nimplicit world prior, enriching the photo supervision for 3D reconstruction. By\nleveraging the proposed method, we effectively augment existing datasets for\nneural rendering methods. Our experimental results indicate that our method\nimproves the reconstruction performance on both static and dynamic scenes."}
+{"qid": "2505.22458v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nUnsupervised domain adaptation for semantic segmentation (UDA-SS) aims to\ntransfer knowledge from labeled source data to unlabeled target data. However,\ntraditional UDA-SS methods assume that category settings between source and\ntarget domains are known, which is unrealistic in real-world scenarios. This\nleads to performance degradation if private classes exist. To address this\nlimitation, we propose Universal Domain Adaptation for Semantic Segmentation\n(UniDA-SS), achieving robust adaptation even without prior knowledge of\ncategory settings. We define the problem in the UniDA-SS scenario as low\nconfidence scores of common classes in the target domain, which leads to\nconfusion with private classes. To solve this problem, we propose UniMAP:\nUniDA-SS with Image Matching and Prototype-based Distinction, a novel framework\ncomposed of two key components. First, Domain-Specific Prototype-based\nDistinction (DSPD) divides each class into two domain-specific prototypes,\nenabling finer separation of domain-specific features and enhancing the\nidentification of common classes across domains. Second, Target-based Image\nMatching (TIM) selects a source image containing the most common-class pixels\nbased on the target pseudo-label and pairs it in a batch to promote effective\nlearning of common classes. We also introduce a new UniDA-SS benchmark and\ndemonstrate through various experiments that UniMAP significantly outperforms\nbaselines. The code is available at\n\\href{https://github.com/KU-VGI/UniMAP}{this https URL}."}
+{"qid": "2505.22427v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nThis paper presents a groundbreaking approach - the first online automatic\ngeometric calibration method for radar and camera systems. Given the\nsignificant data sparsity and measurement uncertainty in radar height data,\nachieving automatic calibration during system operation has long been a\nchallenge. To address the sparsity issue, we propose a Dual-Perspective\nrepresentation that gathers features from both frontal and bird's-eye views.\nThe frontal view contains rich but sensitive height information, whereas the\nbird's-eye view provides robust features against height uncertainty. We thereby\npropose a novel Selective Fusion Mechanism to identify and fuse reliable\nfeatures from both perspectives, reducing the effect of height uncertainty.\nMoreover, for each view, we incorporate a Multi-Modal Cross-Attention Mechanism\nto explicitly find location correspondences through cross-modal matching.\nDuring the training phase, we also design a Noise-Resistant Matcher to provide\nbetter supervision and enhance the robustness of the matching mechanism against\nsparsity and height uncertainty. Our experimental results, tested on the\nnuScenes dataset, demonstrate that our method significantly outperforms\nprevious radar-camera auto-calibration methods, as well as existing\nstate-of-the-art LiDAR-camera calibration techniques, establishing a new\nbenchmark for future research. The code is available at\nhttps://github.com/nycu-acm/RC-AutoCalib."}
+{"qid": "2505.22167v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDiffusion transformers (DiT) have demonstrated exceptional performance in\nvideo generation. However, their large number of parameters and high\ncomputational complexity limit their deployment on edge devices. Quantization\ncan reduce storage requirements and accelerate inference by lowering the\nbit-width of model parameters. Yet, existing quantization methods for image\ngeneration models do not generalize well to video generation tasks. We identify\ntwo primary challenges: the loss of information during quantization and the\nmisalignment between optimization objectives and the unique requirements of\nvideo generation. To address these challenges, we present Q-VDiT, a\nquantization framework specifically designed for video DiT models. From the\nquantization perspective, we propose the Token-aware Quantization Estimator\n(TQE), which compensates for quantization errors in both the token and feature\ndimensions. From the optimization perspective, we introduce Temporal\nMaintenance Distillation (TMD), which preserves the spatiotemporal correlations\nbetween frames and enables the optimization of each frame with respect to the\noverall video context. Our W3A6 Q-VDiT achieves a scene consistency of 23.40,\nsetting a new benchmark and outperforming current state-of-the-art quantization\nmethods by 1.9$\\times$. Code will be available at\nhttps://github.com/cantbebetter2/Q-VDiT."}
+{"qid": "2505.22552v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIntegrating knowledge graphs (KGs) to enhance the reasoning capabilities of\nlarge language models (LLMs) is an emerging research challenge in claim\nverification. While KGs provide structured, semantically rich representations\nwell-suited for reasoning, most existing verification methods rely on\nunstructured text corpora, limiting their ability to effectively leverage KGs.\nAdditionally, despite possessing strong reasoning abilities, modern LLMs\nstruggle with multi-step modular pipelines and reasoning over KGs without\nadaptation. To address these challenges, we propose ClaimPKG, an end-to-end\nframework that seamlessly integrates LLM reasoning with structured knowledge\nfrom KGs. Specifically, the main idea of ClaimPKG is to employ a lightweight,\nspecialized LLM to represent the input claim as pseudo-subgraphs, guiding a\ndedicated subgraph retrieval module to identify relevant KG subgraphs. These\nretrieved subgraphs are then processed by a general-purpose LLM to produce the\nfinal verdict and justification. Extensive experiments on the FactKG dataset\ndemonstrate that ClaimPKG achieves state-of-the-art performance, outperforming\nstrong baselines in this research field by 9%-12% accuracy points across\nmultiple categories. Furthermore, ClaimPKG exhibits zero-shot generalizability\nto unstructured datasets such as HoVer and FEVEROUS, effectively combining\nstructured knowledge from KGs with LLM reasoning across various LLM backbones."}
+{"qid": "2504.21752v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDespite differential privacy (DP) often being considered the de facto\nstandard for data privacy, its realization is vulnerable to unfaithful\nexecution of its mechanisms by servers, especially in distributed settings.\nSpecifically, servers may sample noise from incorrect distributions or generate\ncorrelated noise while appearing to follow established protocols. This work\nanalyzes these malicious behaviors in a general differential privacy framework\nwithin a distributed client-server-verifier setup. To address these adversarial\nproblems, we propose a novel definition called Verifiable Distributed\nDifferential Privacy (VDDP) by incorporating additional verification\nmechanisms. We also explore the relationship between zero-knowledge proofs\n(ZKP) and DP, demonstrating that while ZKPs are sufficient for achieving DP\nunder verifiability requirements, they are not necessary. Furthermore, we\ndevelop two novel and efficient mechanisms that satisfy VDDP: (1) the\nVerifiable Distributed Discrete Laplacian Mechanism (VDDLM), which offers up to\na $4 \\times 10^5$x improvement in proof generation efficiency with only\n0.1-0.2x error compared to the previous state-of-the-art verifiable\ndifferentially private mechanism; (2) an improved solution to Verifiable\nRandomized Response (VRR) under local DP, a special case of VDDP, achieving up\na reduction of up to 5000x in communication costs and the verifier's overhead."}
+{"qid": "2504.21282v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nNatural language (NL)-driven table discovery identifies relevant tables from\nlarge table repositories based on NL queries. While current deep-learning-based\nmethods using the traditional dense vector search pipeline, i.e.,\nrepresentation-index-search, achieve remarkable accuracy, they face several\nlimitations that impede further performance improvements: (i) the errors\naccumulated during the table representation and indexing phases affect the\nsubsequent search accuracy; and (ii) insufficient query-table interaction\nhinders effective semantic alignment, impeding accuracy improvements. In this\npaper, we propose a novel framework Birdie, using a differentiable search\nindex. It unifies the indexing and search into a single encoder-decoder\nlanguage model, thus getting rid of error accumulations. Birdie first assigns\neach table a prefix-aware identifier and leverages a large language model-based\nquery generator to create synthetic queries for each table. It then encodes the\nmapping between synthetic queries/tables and their corresponding table\nidentifiers into the parameters of an encoder-decoder language model, enabling\ndeep query-table interactions. During search, the trained model directly\ngenerates table identifiers for a given query. To accommodate the continual\nindexing of dynamic tables, we introduce an index update strategy via parameter\nisolation, which mitigates the issue of catastrophic forgetting. Extensive\nexperiments demonstrate that Birdie outperforms state-of-the-art dense methods\nby 16.8% in accuracy, and reduces forgetting by over 90% compared to other\ncontinual learning approaches."}
+{"qid": "2504.17448v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nActive learning (AL) reduces human annotation costs for machine learning\nsystems by strategically selecting the most informative unlabeled data for\nannotation, but performing it individually may still be insufficient due to\nrestricted data diversity and annotation budget. Federated Active Learning\n(FAL) addresses this by facilitating collaborative data selection and model\ntraining, while preserving the confidentiality of raw data samples. Yet,\nexisting FAL methods fail to account for the heterogeneity of data distribution\nacross clients and the associated fluctuations in global and local model\nparameters, adversely affecting model accuracy. To overcome these challenges,\nwe propose CHASe (Client Heterogeneity-Aware Data Selection), specifically\ndesigned for FAL. CHASe focuses on identifying those unlabeled samples with\nhigh epistemic variations (EVs), which notably oscillate around the decision\nboundaries during training. To achieve both effectiveness and efficiency,\n\\model{} encompasses techniques for 1) tracking EVs by analyzing inference\ninconsistencies across training epochs, 2) calibrating decision boundaries of\ninaccurate models with a new alignment loss, and 3) enhancing data selection\nefficiency via a data freeze and awaken mechanism with subset sampling.\nExperiments show that CHASe surpasses various established baselines in terms of\neffectiveness and efficiency, validated across diverse datasets, model\ncomplexities, and heterogeneous federation settings."}
+{"qid": "2504.14861v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nMaximum Inner Product Search (MIPS) is a fundamental challenge in machine\nlearning and information retrieval, particularly in high-dimensional data\napplications. Existing approaches to MIPS either rely solely on Inner Product\n(IP) similarity, which faces issues with local optima and redundant\ncomputations, or reduce the MIPS problem to the Nearest Neighbor Search under\nthe Euclidean metric via space projection, leading to topology destruction and\ninformation loss. Despite the divergence of the two paradigms, we argue that\nthere is no inherent binary opposition between IP and Euclidean metrics. By\nstitching IP and Euclidean in the design of indexing and search algorithms, we\ncan significantly enhance MIPS performance. Specifically, this paper explores\nthe theoretical and empirical connections between these two metrics from the\nMIPS perspective. Our investigation, grounded in graph-based search, reveals\nthat different indexing and search strategies offer distinct advantages for\nMIPS, depending on the underlying data topology. Building on these insights, we\nintroduce a novel graph-based index called Metric-Amphibious Graph (MAG) and a\ncorresponding search algorithm, Adaptive Navigation with Metric Switch (ANMS).\nTo facilitate parameter tuning for optimal performance, we identify three\nstatistical indicators that capture essential data topology properties and\ncorrelate strongly with parameter tuning. Extensive experiments on 12\nreal-world datasets demonstrate that MAG outperforms existing state-of-the-art\nmethods, achieving up to 4x search speedup while maintaining adaptability and\nscalability."}
+{"qid": "2504.06975v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nIn order to achieve low latency, high throughput, and partition tolerance,\nmodern databases forgo strong transaction isolation for weak isolation\nguarantees. However, several production databases have been found to suffer\nfrom isolation bugs, breaking their data-consistency contract. Black-box\ntesting is a prominent technique for detecting isolation bugs, by checking\nwhether histories of database transactions adhere to a prescribed isolation\nlevel.\n  Testing databases on realistic workloads of large size requires isolation\ntesters to be as efficient as possible, a requirement that has initiated a\nstudy of the complexity of isolation testing. Although testing strong isolation\nhas been known to be NP-complete, weak isolation levels were recently shown to\nbe testable in polynomial time, which has propelled the scalability of testing\ntools. However, existing testers have a large polynomial complexity,\nrestricting testing to workloads of only moderate size, which is not typical of\nlarge-scale databases.\n  In this work, we develop AWDIT, a highly-efficient and provably optimal\ntester for weak database isolation. Given a history $H$ of size $n$ and $k$\nsessions, AWDIT tests whether H satisfies the most common weak isolation levels\nof Read Committed (RC), Read Atomic (RA), and Causal Consistency (CC) in time\n$O(n^{3/2})$, $O(n^{3/2})$, and $O(n \\cdot k)$, respectively, improving\nsignificantly over the state of the art. Moreover, we prove that AWDIT is\nessentially optimal, in the sense that there is a conditional lower bound of\n$n^{3/2}$ for any weak isolation level between RC and CC. Our experiments show\nthat AWDIT is significantly faster than existing, highly optimized testers;\ne.g., for the $\\sim$20% largest histories, AWDIT obtains an average speedup of\n$245\\times$, $193\\times$, and $62\\times$ for RC, RA, and CC, respectively, over\nthe best baseline."}
+{"qid": "2506.01833v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nInspired by the success of unsupervised pre-training paradigms, researchers\nhave applied these approaches to DNA pre-training. However, we argue that these\napproaches alone yield suboptimal results because pure DNA sequences lack\nsufficient information, since their functions are regulated by genomic profiles\nlike chromatin accessibility. Here, we demonstrate that supervised training for\ngenomic profile prediction serves as a more effective alternative to pure\nsequence pre-training. Furthermore, considering the multi-species and\nmulti-profile nature of genomic profile prediction, we introduce our\n$\\textbf{S}$pecies-$\\textbf{P}$rofile $\\textbf{A}$daptive\n$\\textbf{C}$ollaborative $\\textbf{E}$xperts (SPACE) that leverages Mixture of\nExperts (MoE) to better capture the relationships between DNA sequences across\ndifferent species and genomic profiles, thereby learning more effective DNA\nrepresentations. Through extensive experiments across various tasks, our model\nachieves state-of-the-art performance, establishing that DNA models trained\nwith supervised genomic profiles serve as powerful DNA representation learners.\nThe code is available at https://github.com/ZhuJiwei111/SPACE."}
+{"qid": "2506.00382v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nUnderstanding how feature representations evolve across layers in large\nlanguage models (LLMs) is key to improving their interpretability and\nrobustness. While recent studies have identified critical layers linked to\nspecific functions or behaviors, these efforts typically rely on data-dependent\nanalyses of fine-tuned models, limiting their use to post-hoc settings. In\ncontrast, we introduce a data-oblivious approach to identify intrinsic critical\nlayers in pre-fine-tuned LLMs by analyzing representation dynamics via Centered\nKernel Alignment(CKA). We show that layers with significant shifts in\nrepresentation space are also those most affected during fine-tuning--a pattern\nthat holds consistently across tasks for a given model. Our spectral analysis\nfurther reveals that these shifts are driven by changes in the top principal\ncomponents, which encode semantic transitions from rationales to conclusions.\nWe further apply these findings to two practical scenarios: efficient domain\nadaptation, where fine-tuning critical layers leads to greater loss reduction\ncompared to non-critical layers; and backdoor defense, where freezing them\nreduces attack success rates by up to 40%."}
+{"qid": "2506.00205v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nRehearsal-based methods have shown superior performance in addressing\ncatastrophic forgetting in continual learning (CL) by storing and training on a\nsubset of past data alongside new data in current task. While such a concurrent\nrehearsal strategy is widely used, it remains unclear if this approach is\nalways optimal. Inspired by human learning, where sequentially revisiting tasks\nhelps mitigate forgetting, we explore whether sequential rehearsal can offer\ngreater benefits for CL compared to standard concurrent rehearsal. To address\nthis question, we conduct a theoretical analysis of rehearsal-based CL in\noverparameterized linear models, comparing two strategies: 1) Concurrent\nRehearsal, where past and new data are trained together, and 2) Sequential\nRehearsal, where new data is trained first, followed by revisiting past data\nsequentially. By explicitly characterizing forgetting and generalization error,\nwe show that sequential rehearsal performs better when tasks are less similar.\nThese insights further motivate a novel Hybrid Rehearsal method, which trains\nsimilar tasks concurrently and revisits dissimilar tasks sequentially. We\ncharacterize its forgetting and generalization performance, and our experiments\nwith deep neural networks further confirm that the hybrid approach outperforms\nstandard concurrent rehearsal. This work provides the first comprehensive\ntheoretical analysis of rehearsal-based CL."}
+{"qid": "2505.24835v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nFund allocation has been an increasingly important problem in the financial\ndomain. In reality, we aim to allocate the funds to buy certain assets within a\ncertain future period. Naive solutions such as prediction-only or\nPredict-then-Optimize approaches suffer from goal mismatch. Additionally, the\nintroduction of the SOTA time series forecasting model inevitably introduces\nadditional uncertainty in the predicted result. To solve both problems\nmentioned above, we introduce a Risk-aware Time-Series Predict-and-Allocate\n(RTS-PnO) framework, which holds no prior assumption on the forecasting models.\nSuch a framework contains three features: (i) end-to-end training with\nobjective alignment measurement, (ii) adaptive forecasting uncertainty\ncalibration, and (iii) agnostic towards forecasting models. The evaluation of\nRTS-PnO is conducted over both online and offline experiments. For offline\nexperiments, eight datasets from three categories of financial applications are\nused: Currency, Stock, and Cryptos. RTS-PnO consistently outperforms other\ncompetitive baselines. The online experiment is conducted on the Cross-Border\nPayment business at FiT, Tencent, and an 8.4\\% decrease in regret is witnessed\nwhen compared with the product-line approach. The code for the offline\nexperiment is available at https://github.com/fuyuanlyu/RTS-PnO."}
+{"qid": "2505.24203v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nProtein dynamics play a crucial role in protein biological functions and\nproperties, and their traditional study typically relies on time-consuming\nmolecular dynamics (MD) simulations conducted in silico. Recent advances in\ngenerative modeling, particularly denoising diffusion models, have enabled\nefficient accurate protein structure prediction and conformation sampling by\nlearning distributions over crystallographic structures. However, effectively\nintegrating physical supervision into these data-driven approaches remains\nchallenging, as standard energy-based objectives often lead to intractable\noptimization. In this paper, we introduce Energy-based Alignment (EBA), a\nmethod that aligns generative models with feedback from physical models,\nefficiently calibrating them to appropriately balance conformational states\nbased on their energy differences. Experimental results on the MD ensemble\nbenchmark demonstrate that EBA achieves state-of-the-art performance in\ngenerating high-quality protein ensembles. By improving the physical\nplausibility of generated structures, our approach enhances model predictions\nand holds promise for applications in structural biology and drug discovery."}
+{"qid": "2506.02847v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDeploying large language models (LLMs) on edge devices is crucial for\ndelivering fast responses and ensuring data privacy. However, the limited\nstorage, weight, and power of edge devices make it difficult to deploy\nLLM-powered applications. These devices must balance latency requirements with\nenergy consumption and model accuracy. In this paper, we first quantify the\nchallenges of deploying LLMs on off-the-shelf edge devices and then we present\nCLONE, an in-depth algorithm-hardware co-design at both the model- and\nsystem-level that intelligently integrates real-time, energy optimization while\nmaintaining robust generality. In order to maximize the synergistic benefits of\nthese algorithms in always-on and intermediate edge computing settings, we\nspecialize in a 28nm scalable hardware accelerator system. We implement and\nextensively evaluate CLONE on two off-the-shelf edge platforms. Experiments\nshow that CLONE effectively accelerates the inference process up to 11.92x, and\nsaves energy up to 7.36x, while maintaining high-generation."}
+{"qid": "2505.22194v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nVision Transformers (ViTs) leverage the transformer architecture to\neffectively capture global context, demonstrating strong performance in\ncomputer vision tasks. A major challenge in ViT hardware acceleration is that\nthe model family contains complex arithmetic operations that are sensitive to\nmodel accuracy, such as the Softmax and LayerNorm operations, which cannot be\nmapped onto efficient hardware with low precision. Existing methods only\nexploit parallelism in the matrix multiplication operations of the model on\nhardware and keep these complex operations on the CPU. This results in\nsuboptimal performance due to the communication overhead between the CPU and\naccelerator. Can new data formats solve this problem?\n  In this work, we present the first ViT accelerator that maps all operations\nof the ViT models onto FPGAs. We exploit a new arithmetic format named\nMicroscaling Integer (MXInt) for datapath designs and evaluate how different\ndesign choices can be made to trade off accuracy, hardware performance, and\nhardware utilization. Our contributions are twofold. First, we quantize ViTs\nusing the MXInt format, achieving both high area efficiency and accuracy.\nSecond, we propose MXInt-specific hardware optimization that map these complex\narithmetic operations into custom hardware. Within 1\\% accuracy loss, our\nmethod achieves at least 93$\\times$ speedup compared to Float16 and at least\n1.9$\\times$ speedup compared to related work."}
+{"qid": "2505.11554v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nMemory bandwidth regulation and cache partitioning are widely used techniques\nfor achieving predictable timing in real-time computing systems. Combined with\npartitioned scheduling, these methods require careful co-allocation of tasks\nand resources to cores, as task execution times strongly depend on available\nallocated resources. To address this challenge, this paper presents a 0-1\nlinear program for task-resource co-allocation, along with a multi-objective\nheuristic designed to minimize resource usage while guaranteeing schedulability\nunder a preemptive EDF scheduling policy. Our heuristic employs a multi-layer\nframework, where an outer layer explores resource allocations using\nPareto-pruned search, and an inner layer optimizes task allocation by solving a\nknapsack problem using dynamic programming. To evaluate the performance of the\nproposed optimization algorithm, we profile real-world benchmarks on an\nembedded AMD UltraScale+ ZCU102 platform, with fine-grained resource\npartitioning enabled by the Jailhouse hypervisor, leveraging cache set\npartitioning and MemGuard for memory bandwidth regulation. Experiments based on\nthe benchmarking results show that the proposed 0-1 linear program outperforms\nexisting mixed-integer programs by finding more optimal solutions within the\nsame time limit. Moreover, the proposed multi-objective multi-layer heuristic\nperforms consistently better than the state-of-the-art multi-resource-task\nco-allocation algorithm in terms of schedulability, resource usage, number of\nnon-dominated solutions, and computational efficiency."}
+{"qid": "2505.08071v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDe novo assembly enables investigations of unknown genomes, paving the way\nfor personalized medicine and disease management. However, it faces immense\ncomputational challenges arising from the excessive data volumes and\nalgorithmic complexity.\n  While state-of-the-art de novo assemblers utilize distributed systems for\nextreme-scale genome assembly, they demand substantial computational and memory\nresources. They also fail to address the inherent challenges of de novo\nassembly, including a large memory footprint, memory-bound behavior, and\nirregular data patterns stemming from complex, interdependent data structures.\nGiven these challenges, de novo assembly merits a custom hardware solution,\nthough existing approaches have not fully addressed the limitations.\n  We propose NMP-PaK, a hardware-software co-design that accelerates scalable\nde novo genome assembly through near-memory processing (NMP). Our channel-level\nNMP architecture addresses memory bottlenecks while providing sufficient\nscratchpad space for processing elements. Customized processing elements\nmaximize parallelism while efficiently handling large data structures that are\nboth dynamic and interdependent. Software optimizations include customized\nbatch processing to reduce the memory footprint and hybrid CPU-NMP processing\nto address hardware underutilization caused by irregular data patterns.\n  NMP-PaK conducts the same genome assembly while incurring a 14X smaller\nmemory footprint compared to the state-of-the-art de novo assembly. Moreover,\nNMP-PaK delivers a 16X performance improvement over the CPU baseline, with a\n2.4X reduction in memory operations. Consequently, NMP-PaK achieves 8.3X\ngreater throughput than state-of-the-art de novo assembly under the same\nresource constraints, showcasing its superior computational efficiency."}
+{"qid": "2504.06211v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nZero-Knowledge Proofs (ZKPs) are rapidly gaining importance in\nprivacy-preserving and verifiable computing. ZKPs enable a proving party to\nprove the truth of a statement to a verifying party without revealing anything\nelse. ZKPs have applications in blockchain technologies, verifiable machine\nlearning, and electronic voting, but have yet to see widespread adoption due to\nthe computational complexity of the proving process. Recent works have\naccelerated the key primitives of state-of-the-art ZKP protocols on GPU and\nASIC. However, the protocols accelerated thus far face one of two challenges:\nthey either require a trusted setup for each application, or they generate\nlarger proof sizes with higher verification costs, limiting their applicability\nin scenarios with numerous verifiers or strict verification time constraints.\nThis work presents an accelerator, zkSpeed, for HyperPlonk, a state-of-the-art\nZKP protocol that supports both one-time, universal setup and small proof sizes\nfor typical ZKP applications in publicly verifiable, consensus-based systems.\nWe accelerate the entire protocol, including two major primitives: SumCheck and\nMulti-scalar Multiplications (MSMs). We develop a full-chip architecture using\n366.46 mm$^2$ and 2 TB/s of bandwidth to accelerate the entire proof generation\nprocess, achieving geometric mean speedups of 801$\\times$ over CPU baselines."}
+{"qid": "2504.19283v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nServerless computing abstracts away server management, enabling automatic\nscaling, efficient resource utilization, and cost-effective pricing models.\nHowever, despite these advantages, it faces the significant challenge of\ncold-start latency, adversely impacting end-to-end performance. Our study shows\nthat many serverless functions initialize libraries that are rarely or never\nused under typical workloads, thus introducing unnecessary overhead. Although\nexisting static analysis techniques can identify unreachable libraries, they\nfail to address workload-dependent inefficiencies, resulting in limited\nperformance improvements. To overcome these limitations, we present SLIMSTART,\na profile-guided optimization tool designed to identify and mitigate\ninefficient library usage patterns in serverless applications. By leveraging\nstatistical sampling and call-path profiling, SLIMSTART collects runtime\nlibrary usage data, generates detailed optimization reports, and applies\nautomated code transformations to reduce cold-start overhead. Furthermore,\nSLIMSTART integrates seamlessly into CI/CD pipelines, enabling adaptive\nmonitoring and continuous optimizations tailored to evolving workloads. Through\nextensive evaluation across three benchmark suites and four real-world\nserverless applications, SLIMSTART achieves up to a 2.30X speedup in\ninitialization latency, a 2.26X improvement in end-to-end latency, and a 1.51X\nreduction in memory usage, demonstrating its effectiveness in addressing\ncold-start inefficiencies and optimizing resource utilization."}
+{"qid": "2504.11007v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nModern cloud-native applications increasingly utilise managed cloud services\nand containerisation technologies, such as Kubernetes, to achieve rapid\ntime-to-market and scalable deployments. Organisations must consider various\nfactors, including cost implications when deciding on a hosting platform for\ncontainerised applications as the usage grows. An emerging discipline called\nFinOps combines financial management and cloud operations to optimise costs in\ncloud-based applications. While prior research has explored system-level\noptimisation strategies for cost and resource efficiency in containerized\nsystems, analysing network costs in Kubernetes clusters remains underexplored.\nThis paper investigates the network usage and cost implications of\ncontainerised applications running on Kubernetes clusters. Using a methodology\nthat combines measurement analysis, experimentation, and cost modelling, we aim\nto provide organisations with actionable insights into network cost\noptimisation. Our findings highlight key considerations for analysing network\nexpenditures and evaluating the potential cost benefits of deploying\napplications on cloud providers. Overall, this paper contributes to the\nemerging FinOps discipline by addressing the financial and operational aspects\nof managing network costs in cloud-native environments."}
+{"qid": "2504.09307v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nTraining LLMs in distributed environments presents significant challenges due\nto the complexity of model execution, deployment systems, and the vast space of\nconfigurable strategies. Although various optimization techniques exist,\nachieving high efficiency in practice remains difficult. Accurate performance\nmodels that effectively characterize and predict a model's behavior are\nessential for guiding optimization efforts and system-level studies. We propose\nLumos, a trace-driven performance modeling and estimation toolkit for\nlarge-scale LLM training, designed to accurately capture and predict the\nexecution behaviors of modern LLMs. We evaluate Lumos on a production ML\ncluster with up to 512 NVIDIA H100 GPUs using various GPT-3 variants,\ndemonstrating that it can replay execution time with an average error of just\n3.3%, along with other runtime details, across different models and\nconfigurations. Additionally, we validate its ability to estimate performance\nfor new setups from existing traces, facilitating efficient exploration of\nmodel and deployment configurations."}
+{"qid": "2506.02750v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nLearning vectorized embeddings is fundamental to many recommender systems for\nuser-item matching. To enable efficient online inference, representation\nbinarization, which embeds latent features into compact binary sequences, has\nrecently shown significant promise in optimizing both memory usage and\ncomputational overhead. However, existing approaches primarily focus on\nnumerical quantization, neglecting the associated information loss, which often\nresults in noticeable performance degradation. To address these issues, we\nstudy the problem of graph representation binarization for efficient\ncollaborative filtering. Our findings indicate that explicitly mitigating\ninformation loss at various stages of embedding binarization has a significant\npositive impact on performance. Building on these insights, we propose an\nenhanced framework, BiGeaR++, which specifically leverages supervisory signals\nfrom pseudo-positive samples, incorporating both real item data and latent\nembedding samples. Compared to its predecessor BiGeaR, BiGeaR++ introduces a\nfine-grained inference distillation mechanism and an effective embedding sample\nsynthesis approach. Empirical evaluations across five real-world datasets\ndemonstrate that the new designs in BiGeaR++ work seamlessly well with other\nmodules, delivering substantial improvements of around 1%-10% over BiGeaR and\nthus achieving state-of-the-art performance compared to the competing methods.\nOur implementation is available at https://github.com/QueYork/BiGeaR-SS."}
+{"qid": "2505.23452v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nOpinion mining plays a vital role in analysing user feedback and extracting\ninsights from textual data. While most research focuses on sentiment polarity\n(e.g., positive, negative, neutral), fine-grained emotion classification in app\nreviews remains underexplored. This paper addresses this gap by identifying and\naddressing the challenges and limitations in fine-grained emotion analysis in\nthe context of app reviews. Our study adapts Plutchik's emotion taxonomy to app\nreviews by developing a structured annotation framework and dataset. Through an\niterative human annotation process, we define clear annotation guidelines and\ndocument key challenges in emotion classification. Additionally, we evaluate\nthe feasibility of automating emotion annotation using large language models,\nassessing their cost-effectiveness and agreement with human-labelled data. Our\nfindings reveal that while large language models significantly reduce manual\neffort and maintain substantial agreement with human annotators, full\nautomation remains challenging due to the complexity of emotional\ninterpretation. This work contributes to opinion mining by providing structured\nguidelines, an annotated dataset, and insights for developing automated\npipelines to capture the complexity of emotions in app reviews."}
+{"qid": "2505.21811v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nSequential recommendation is a popular paradigm in modern recommender\nsystems. In particular, one challenging problem in this space is cross-domain\nsequential recommendation (CDSR), which aims to predict future behaviors given\nuser interactions across multiple domains. Existing CDSR frameworks are mostly\nbuilt on the self-attention transformer and seek to improve by explicitly\ninjecting additional domain-specific components (e.g. domain-aware module\nblocks). While these additional components help, we argue they overlook the\ncore self-attention module already present in the transformer, a naturally\npowerful tool to learn correlations among behaviors. In this work, we aim to\nimprove the CDSR performance for simple models from a novel perspective of\nenhancing the self-attention. Specifically, we introduce a Pareto-optimal\nself-attention and formulate the cross-domain learning as a multi-objective\nproblem, where we optimize the recommendation task while dynamically minimizing\nthe cross-domain attention scores. Our approach automates knowledge transfer in\nCDSR (dubbed as AutoCDSR) -- it not only mitigates negative transfer but also\nencourages complementary knowledge exchange among auxiliary domains. Based on\nthe idea, we further introduce AutoCDSR+, a more performant variant with slight\nadditional cost. Our proposal is easy to implement and works as a plug-and-play\nmodule that can be incorporated into existing transformer-based recommenders.\nBesides flexibility, it is practical to deploy because it brings little extra\ncomputational overheads without heavy hyper-parameter tuning. AutoCDSR on\naverage improves Recall@10 for SASRec and Bert4Rec by 9.8% and 16.0% and\nNDCG@10 by 12.0% and 16.7%, respectively. Code is available at\nhttps://github.com/snap-research/AutoCDSR."}
+{"qid": "2505.20227v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nMulti-Domain Recommendation (MDR) achieves the desirable recommendation\nperformance by effectively utilizing the transfer information across different\ndomains. Despite the great success, most existing MDR methods adopt a single\nstructure to transfer complex domain-shared knowledge. However, the beneficial\ntransferring information should vary across different domains. When there is\nknowledge conflict between domains or a domain is of poor quality,\nunselectively leveraging information from all domains will lead to a serious\nNegative Transfer Problem (NTP). Therefore, how to effectively model the\ncomplex transfer relationships between domains to avoid NTP is still a\ndirection worth exploring. To address these issues, we propose a simple and\ndynamic Similar Domain Selection Principle (SDSP) for multi-domain\nrecommendation in this paper. SDSP presents the initial exploration of\nselecting suitable domain knowledge for each domain to alleviate NTP.\nSpecifically, we propose a novel prototype-based domain distance measure to\neffectively model the complexity relationship between domains. Thereafter, the\nproposed SDSP can dynamically find similar domains for each domain based on the\nsupervised signals of the domain metrics and the unsupervised distance measure\nfrom the learned domain prototype. We emphasize that SDSP is a lightweight\nmethod that can be incorporated with existing MDR methods for better\nperformance while not introducing excessive time overheads. To the best of our\nknowledge, it is the first solution that can explicitly measure domain-level\ngaps and dynamically select appropriate domains in the MDR field. Extensive\nexperiments on three datasets demonstrate the effectiveness of our proposed\nmethod."}
+{"qid": "2505.19356v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nNeural retrieval methods using transformer-based pre-trained language models\nhave advanced multilingual and cross-lingual retrieval. However, their\neffectiveness for low-resource, morphologically rich languages such as Amharic\nremains underexplored due to data scarcity and suboptimal tokenization. We\naddress this gap by introducing Amharic-specific dense retrieval models based\non pre-trained Amharic BERT and RoBERTa backbones. Our proposed\nRoBERTa-Base-Amharic-Embed model (110M parameters) achieves a 17.6% relative\nimprovement in MRR@10 and a 9.86% gain in Recall@10 over the strongest\nmultilingual baseline, Arctic Embed 2.0 (568M parameters). More compact\nvariants, such as RoBERTa-Medium-Amharic-Embed (42M), remain competitive while\nbeing over 13x smaller. Additionally, we train a ColBERT-based late interaction\nretrieval model that achieves the highest MRR@10 score (0.843) among all\nevaluated models. We benchmark our proposed models against both sparse and\ndense retrieval baselines to systematically assess retrieval effectiveness in\nAmharic. Our analysis highlights key challenges in low-resource settings and\nunderscores the importance of language-specific adaptation. To foster future\nresearch in low-resource IR, we publicly release our dataset, codebase, and\ntrained models at https://github.com/kidist-amde/amharic-ir-benchmarks."}
+{"qid": "2505.19307v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nNeural retrieval models excel in Web search, but their training requires\nsubstantial amounts of labeled query-document pairs, which are costly to\nobtain. With the widespread availability of Web document collections like\nClueWeb22, synthetic queries generated by large language models offer a\nscalable alternative. Still, synthetic training queries often vary in quality,\nwhich leads to suboptimal downstream retrieval performance. Existing methods\ntypically filter out noisy query-document pairs based on signals from an\nexternal re-ranker. In contrast, we propose a framework that leverages Direct\nPreference Optimization (DPO) to integrate ranking signals into the query\ngeneration process, aiming to directly optimize the model towards generating\nhigh-quality queries that maximize downstream retrieval effectiveness.\nExperiments show higher ranker-assessed relevance between query-document pairs\nafter DPO, leading to stronger downstream performance on the MS~MARCO benchmark\nwhen compared to baseline models trained with synthetic data."}
+{"qid": "2505.17507v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nThe rapid growth of open source machine learning (ML) resources, such as\nmodels and datasets, has accelerated IR research. However, existing platforms\nlike Hugging Face do not explicitly utilize structured representations,\nlimiting advanced queries and analyses such as tracing model evolution and\nrecommending relevant datasets. To fill the gap, we construct HuggingKG, the\nfirst large-scale knowledge graph built from the Hugging Face community for ML\nresource management. With 2.6 million nodes and 6.2 million edges, HuggingKG\ncaptures domain-specific relations and rich textual attributes. It enables us\nto further present HuggingBench, a multi-task benchmark with three novel test\ncollections for IR tasks including resource recommendation, classification, and\ntracing. Our experiments reveal unique characteristics of HuggingKG and the\nderived tasks. Both resources are publicly available, expected to advance\nresearch in open source resource sharing and management."}
+{"qid": "2505.12791v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nThis paper reports on findings from a comparative study on the effectiveness\nand efficiency of federated unlearning strategies within Federated Online\nLearning to Rank (FOLTR), with specific attention to systematically analysing\nthe unlearning capabilities of methods in a verifiable manner.\n  Federated approaches to ranking of search results have recently garnered\nattention to address users privacy concerns. In FOLTR, privacy is safeguarded\nby collaboratively training ranking models across decentralized data sources,\npreserving individual user data while optimizing search results based on\nimplicit feedback, such as clicks.\n  Recent legislation introduced across numerous countries is establishing the\nso called \"the right to be forgotten\", according to which services based on\nmachine learning models like those in FOLTR should provide capabilities that\nallow users to remove their own data from those used to train models. This has\nsparked the development of unlearning methods, along with evaluation practices\nto measure whether unlearning of a user data successfully occurred. Current\nevaluation practices are however often controversial, necessitating the use of\nmultiple metrics for a more comprehensive assessment -- but previous proposals\nof unlearning methods only used single evaluation metrics.\n  This paper addresses this limitation: our study rigorously assesses the\neffectiveness of unlearning strategies in managing both under-unlearning and\nover-unlearning scenarios using adapted, and newly proposed evaluation metrics.\nThanks to our detailed analysis, we uncover the strengths and limitations of\nfive unlearning strategies, offering valuable insights into optimizing\nfederated unlearning to balance data privacy and system performance within\nFOLTR. We publicly release our code and complete results at\nhttps://github.com/Iris1026/Unlearning-for-FOLTR.git."}
+{"qid": "2505.07166v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nDense retrievers utilize pre-trained backbone language models (e.g., BERT,\nLLaMA) that are fine-tuned via contrastive learning to perform the task of\nencoding text into sense representations that can be then compared via a\nshallow similarity operation, e.g. inner product. Recent research has\nquestioned the role of fine-tuning vs. that of pre-training within dense\nretrievers, specifically arguing that retrieval knowledge is primarily gained\nduring pre-training, meaning knowledge not acquired during pre-training cannot\nbe sub-sequentially acquired via fine-tuning. We revisit this idea here as the\nclaim was only studied in the context of a BERT-based encoder using DPR as\nrepresentative dense retriever. We extend the previous analysis by testing\nother representation approaches (comparing the use of CLS tokens with that of\nmean pooling), backbone architectures (encoder-only BERT vs. decoder-only\nLLaMA), and additional datasets (MSMARCO in addition to Natural Questions). Our\nstudy confirms that in DPR tuning, pre-trained knowledge underpins retrieval\nperformance, with fine-tuning primarily adjusting neuron activation rather than\nreorganizing knowledge. However, this pattern does not hold universally, such\nas in mean-pooled (Contriever) and decoder-based (LLaMA) models. We ensure full\nreproducibility and make our implementation publicly available at\nhttps://github.com/ielab/DenseRetriever-Knowledge-Acquisition."}
+{"qid": "2505.03484v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nRecent deep sequential recommendation models often struggle to effectively\nmodel key characteristics of user behaviors, particularly in handling sequence\nlength variations and capturing diverse interaction patterns. We propose\nSTAR-Rec, a novel architecture that synergistically combines preference-aware\nattention and state-space modeling through a sequence-level mixture-of-experts\nframework. STAR-Rec addresses these challenges by: (1) employing\npreference-aware attention to capture both inherently similar item\nrelationships and diverse preferences, (2) utilizing state-space modeling to\nefficiently process variable-length sequences with linear complexity, and (3)\nincorporating a mixture-of-experts component that adaptively routes different\nbehavioral patterns to specialized experts, handling both focused\ncategory-specific browsing and diverse category exploration patterns. We\ntheoretically demonstrate how the state space model and attention mechanisms\ncan be naturally unified in recommendation scenarios, where SSM captures\ntemporal dynamics through state compression while attention models both similar\nand diverse item relationships. Extensive experiments on four real-world\ndatasets demonstrate that STAR-Rec consistently outperforms state-of-the-art\nsequential recommendation methods, particularly in scenarios involving diverse\nuser behaviors and varying sequence lengths."}
+{"qid": "2505.00552v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nGraph convolutional networks have recently gained prominence in collaborative\nfiltering (CF) for recommendations. However, we identify potential bottlenecks\nin two foundational components. First, the embedding layer leads to a latent\nspace with limited capacity, overlooking locally observed but potentially\nvaluable preference patterns. Also, the widely-used neighborhood aggregation is\nlimited in its ability to leverage diverse preference patterns in a\nfine-grained manner. Building on spectral graph theory, we reveal that these\nlimitations stem from graph filtering with a cut-off in the frequency spectrum\nand a restricted linear form. To address these issues, we introduce ChebyCF, a\nCF framework based on graph spectral filtering. Instead of a learned embedding,\nit takes a user's raw interaction history to utilize the full spectrum of\nsignals contained in it. Also, it adopts Chebyshev interpolation to effectively\napproximate a flexible non-linear graph filter, and further enhances it by\nusing an additional ideal pass filter and degree-based normalization. Through\nextensive experiments, we verify that ChebyCF overcomes the aforementioned\nbottlenecks and achieves state-of-the-art performance across multiple\nbenchmarks and reasonably fast inference. Our code is available at\nhttps://github.com/chanwoo0806/ChebyCF."}
+{"qid": "2504.20458v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nConversational recommendation systems (CRSs) use multi-turn interaction to\ncapture user preferences and provide personalized recommendations. A\nfundamental challenge in CRSs lies in effectively understanding user\npreferences from conversations. User preferences can be multifaceted and\ncomplex, posing significant challenges for accurate recommendations even with\naccess to abundant external knowledge. While interaction with users can clarify\ntheir true preferences, frequent user involvement can lead to a degraded user\nexperience.\n  To address this problem, we propose a generative reward model based simulated\nuser, named GRSU, for automatic interaction with CRSs. The simulated user\nprovides feedback to the items recommended by CRSs, enabling them to better\ncapture intricate user preferences through multi-turn interaction. Inspired by\ngenerative reward models, we design two types of feedback actions for the\nsimulated user: i.e., generative item scoring, which offers coarse-grained\nfeedback, and attribute-based item critique, which provides fine-grained\nfeedback. To ensure seamless integration, these feedback actions are unified\ninto an instruction-based format, allowing the development of a unified\nsimulated user via instruction tuning on synthesized data. With this simulated\nuser, automatic multi-turn interaction with CRSs can be effectively conducted.\nFurthermore, to strike a balance between effectiveness and efficiency, we draw\ninspiration from the paradigm of reward-guided search in complex reasoning\ntasks and employ beam search for the interaction process. On top of this, we\npropose an efficient candidate ranking method to improve the recommendation\nresults derived from interaction. Extensive experiments on public datasets\ndemonstrate the effectiveness, efficiency, and transferability of our approach."}
+{"qid": "2504.18383v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nCross-domain Sequential Recommendation (CDSR) aims to extract the preference\nfrom the user's historical interactions across various domains. Despite some\nprogress in CDSR, two problems set the barrier for further advancements, i.e.,\noverlap dilemma and transition complexity. The former means existing CDSR\nmethods severely rely on users who own interactions on all domains to learn\ncross-domain item relationships, compromising the practicability. The latter\nrefers to the difficulties in learning the complex transition patterns from the\nmixed behavior sequences. With powerful representation and reasoning abilities,\nLarge Language Models (LLMs) are promising to address these two problems by\nbridging the items and capturing the user's preferences from a semantic view.\nTherefore, we propose an LLMs Enhanced Cross-domain Sequential Recommendation\nmodel (LLM4CDSR). To obtain the semantic item relationships, we first propose\nan LLM-based unified representation module to represent items. Then, a\ntrainable adapter with contrastive regularization is designed to adapt the CDSR\ntask. Besides, a hierarchical LLMs profiling module is designed to summarize\nuser cross-domain preferences. Finally, these two modules are integrated into\nthe proposed tri-thread framework to derive recommendations. We have conducted\nextensive experiments on three public cross-domain datasets, validating the\neffectiveness of LLM4CDSR. We have released the code online."}
+{"qid": "2504.17519v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nGenerative retrieval (GR) has emerged as a promising paradigm in information\nretrieval (IR). However, most existing GR models are developed and evaluated\nusing a static document collection, and their performance in dynamic corpora\nwhere document collections evolve continuously is rarely studied. In this\npaper, we first reproduce and systematically evaluate various representative GR\napproaches over dynamic corpora. Through extensive experiments, we reveal that\nexisting GR models with \\textit{text-based} docids show superior generalization\nto unseen documents. We observe that the more fine-grained the docid design in\nthe GR model, the better its performance over dynamic corpora, surpassing BM25\nand even being comparable to dense retrieval methods. While GR models with\n\\textit{numeric-based} docids show high efficiency, their performance drops\nsignificantly over dynamic corpora. Furthermore, our experiments find that the\nunderperformance of numeric-based docids is partly due to their excessive\ntendency toward the initial document set, which likely results from overfitting\non the training set. We then conduct an in-depth analysis of the\nbest-performing GR methods. We identify three critical advantages of text-based\ndocids in dynamic corpora: (i) Semantic alignment with language models'\npretrained knowledge, (ii) Fine-grained docid design, and (iii) High lexical\ndiversity. Building on these insights, we finally propose a novel multi-docid\ndesign that leverages both the efficiency of numeric-based docids and the\neffectiveness of text-based docids, achieving improved performance in dynamic\ncorpus without requiring additional retraining. Our work offers empirical\nevidence for advancing GR methods over dynamic corpora and paves the way for\ndeveloping more generalized yet efficient GR models in real-world search\nengines."}
+{"qid": "2504.15849v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nWith the growing abundance of repositories containing tabular data,\ndiscovering relevant tables for in-depth analysis remains a challenging task.\nExisting table discovery methods primarily retrieve desired tables based on a\nquery table or several vague keywords, leaving users to manually filter large\nresult sets. To address this limitation, we propose a new task: NL-conditional\ntable discovery (nlcTD), where users combine a query table with natural\nlanguage (NL) requirements to refine search results. To advance research in\nthis area, we present nlcTables, a comprehensive benchmark dataset comprising\n627 diverse queries spanning NL-only, union, join, and fuzzy conditions, 22,080\ncandidate tables, and 21,200 relevance annotations. Our evaluation of six\nstate-of-the-art table discovery methods on nlcTables reveals substantial\nperformance gaps, highlighting the need for advanced techniques to tackle this\nchallenging nlcTD scenario. The dataset, construction framework, and baseline\nimplementations are publicly available at\nhttps://github.com/SuDIS-ZJU/nlcTables to foster future research."}
+{"qid": "2504.14991v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nFairness is an increasingly important factor in re-ranking tasks. Prior work\nhas identified a trade-off between ranking accuracy and item fairness. However,\nthe underlying mechanisms are still not fully understood. An analogy can be\ndrawn between re-ranking and the dynamics of economic transactions. The\naccuracy-fairness trade-off parallels the coupling of the commodity tax\ntransfer process. Fairness considerations in re-ranking, similar to a commodity\ntax on suppliers, ultimately translate into a cost passed on to consumers.\nAnalogously, item-side fairness constraints result in a decline in user-side\naccuracy. In economics, the extent to which commodity tax on the supplier (item\nfairness) transfers to commodity tax on users (accuracy loss) is formalized\nusing the notion of elasticity. The re-ranking fairness-accuracy trade-off is\nsimilarly governed by the elasticity of utility between item groups. This\ninsight underscores the limitations of current fair re-ranking evaluations,\nwhich often rely solely on a single fairness metric, hindering comprehensive\nassessment of fair re-ranking algorithms. Centered around the concept of\nelasticity, this work presents two significant contributions. We introduce the\nElastic Fairness Curve (EF-Curve) as an evaluation framework. This framework\nenables a comparative analysis of algorithm performance across different\nelasticity levels, facilitating the selection of the most suitable approach.\nFurthermore, we propose ElasticRank, a fair re-ranking algorithm that employs\nelasticity calculations to adjust inter-item distances within a curved space.\nExperiments on three widely used ranking datasets demonstrate its effectiveness\nand efficiency."}
+{"qid": "2504.14243v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nRanking models primarily focus on modeling the relative order of predictions\nwhile often neglecting the significance of the accuracy of their absolute\nvalues. However, accurate absolute values are essential for certain downstream\ntasks, necessitating the calibration of the original predictions. To address\nthis, existing calibration approaches typically employ predefined\ntransformation functions with order-preserving properties to adjust the\noriginal predictions. Unfortunately, these functions often adhere to fixed\nforms, such as piece-wise linear functions, which exhibit limited\nexpressiveness and flexibility, thereby constraining their effectiveness in\ncomplex calibration scenarios. To mitigate this issue, we propose implementing\na calibrator using an Unconstrained Monotonic Neural Network (UMNN), which can\nlearn arbitrary monotonic functions with great modeling power. This approach\nsignificantly relaxes the constraints on the calibrator, improving its\nflexibility and expressiveness while avoiding excessively distorting the\noriginal predictions by requiring monotonicity. Furthermore, to optimize this\nhighly flexible network for calibration, we introduce a novel additional loss\nfunction termed Smooth Calibration Loss (SCLoss), which aims to fulfill a\nnecessary condition for achieving the ideal calibration state. Extensive\noffline experiments confirm the effectiveness of our method in achieving\nsuperior calibration performance. Moreover, deployment in Kuaishou's\nlarge-scale online video ranking system demonstrates that the method's\ncalibration improvements translate into enhanced business metrics. The source\ncode is available at https://github.com/baiyimeng/UMC."}
+{"qid": "2504.12900v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nPersonalized outfit generation aims to construct a set of compatible and\npersonalized fashion items as an outfit. Recently, generative AI models have\nreceived widespread attention, as they can generate fashion items for users to\ncomplete an incomplete outfit or create a complete outfit. However, they have\nlimitations in terms of lacking diversity and relying on the supervised\nlearning paradigm. Recognizing this gap, we propose a novel framework\nFashionDPO, which fine-tunes the fashion outfit generation model using direct\npreference optimization. This framework aims to provide a general fine-tuning\napproach to fashion generative models, refining a pre-trained fashion outfit\ngeneration model using automatically generated feedback, without the need to\ndesign a task-specific reward function. To make sure that the feedback is\ncomprehensive and objective, we design a multi-expert feedback generation\nmodule which covers three evaluation perspectives, \\ie quality, compatibility\nand personalization. Experiments on two established datasets, \\ie iFashion and\nPolyvore-U, demonstrate the effectiveness of our framework in enhancing the\nmodel's ability to align with users' personalized preferences while adhering to\nfashion compatibility principles. Our code and model checkpoints are available\nat https://github.com/Yzcreator/FashionDPO."}
+{"qid": "2504.09935v1", "query": "Write a Related Works section for an academic paper given the paper's abstract. Here is the paper abstract:\nGenerative retrieval seeks to replace traditional search index data\nstructures with a single large-scale neural network, offering the potential for\nimproved efficiency and seamless integration with generative large language\nmodels. As an end-to-end paradigm, generative retrieval adopts a learned\ndifferentiable search index to conduct retrieval by directly generating\ndocument identifiers through corpus-specific constrained decoding. The\ngeneralization capabilities of generative retrieval on out-of-distribution\ncorpora have gathered significant attention.\n  In this paper, we examine the inherent limitations of constrained\nauto-regressive generation from two essential perspectives: constraints and\nbeam search. We begin with the Bayes-optimal setting where the generative\nretrieval model exactly captures the underlying relevance distribution of all\npossible documents. Then we apply the model to specific corpora by simply\nadding corpus-specific constraints. Our main findings are two-fold: (i) For the\neffect of constraints, we derive a lower bound of the error, in terms of the KL\ndivergence between the ground-truth and the model-predicted step-wise marginal\ndistributions. (ii) For the beam search algorithm used during generation, we\nreveal that the usage of marginal distributions may not be an ideal approach.\nThis paper aims to improve our theoretical understanding of the generalization\ncapabilities of the auto-regressive decoding retrieval paradigm, laying a\nfoundation for its limitations and inspiring future advancements toward more\nrobust and generalizable generative retrieval."}