arXiv subjects

Sukmin Cho

Publications and source records attributed to Sukmin Cho.

At least 19 recordsLinked to original sources

arXiv2026

OasisKV: Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching

Large language model (LLM) inference serving is increasingly constrained by memory rather than compute. As long-context and long-form reasoning workloads become more prevalent, the key-value (KV) cache dominates both memory footprint and memory traffic during LLM token generation, i.e., decode. In particular, HBM capacity has become a scarce and costly resource that heavily limits inference batch size and system throughput. This paper presents OasisKV, a memory-centric LLM inference system design that alleviates HBM capacity pressure by decoupling full KV-cache storage from HBM during LLM decoding. Because decode-time attention is naturally sparse, OasisKV keeps only the KV entries of the most relevant tokens in HBMs for attention computation. We observe that future important tokens can be predicted accurately in advance using lookahead tokens drafted by speculative decoding (SD). OasisKV employs an efficient attention background pipeline to identify important KV blocks. They are then prefetched from higher-capacity memory tiers (e.g., host or remote memory) and staged in HBMs before being used in the next decode step. We implement OasisKV based on vLLM. The lookahead prediction is accurate enough to keep accuracy within 0.7 points of full attention under a 2,048-token KV budget. This lets OasisKV turn sparsity into throughput gain: $1.69\times$ over dense vLLM on the reasoning workload at 0.1 points of accuracy loss, and up to $2.1\times$ on multi-GPU long-context serving. Under prefill--decode disaggregation, OasisKV reaches about $2\times$ dense throughput while admitting each request with $6.5$--$9.7\times$ less KV and holding $2.2$-$2.6$ less decode-node host memory than full KV transfer.

cs.DC↗

arXiv2026

Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval

Sign Language Retrieval (SLRet) enables efficient access to sign language content but remains fragile in fine-grained scenarios where visually similar signs must be distinguished. We show that this limitation does not stem from model capacity, but from ineffective hard negative supervision. Specifically, we formulate fine-grained retrieval failures as a negative distribution mismatch: semantically distinct yet visually confusable signs are rarely treated as hard negatives, while existing text-based mining strategies fail to capture such visual ambiguity. To address this issue, we propose Sign-Aware Hard Negative Mining (SAN), which constructs hard negatives based on visual confusability in the sign embedding space rather than linguistic similarity. Experiments on PHOENIX-2014T demonstrate that SAN substantially improves fine-grained retrieval performance while preserving coarse-grained accuracy, highlighting the importance of aligning negative supervision with visual ambiguity in sign language retrieval.

cs.CV↗

arXiv2026

ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of every distinct expert its batch activates. We present ELDR, an expert-locality-aware decode router for PD-disaggregated MoE serving. From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation. Offline, balanced K-means partitions signature space across decode workers; online, locality-band routing sends each request to the least-loaded worker among those best matching its signature. A signature cache, co-indexed with the KV cache at KV-block granularity, keeps signatures exact under prefix caching. Implemented in vLLM and evaluated on deployments of up to 40 GPUs, ELDR reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads, with model outputs unchanged.

cs.DC↗

arXiv2025

Upcycling Candidate Tokens of Large Language Models for Query Expansion

Query Expansion (QE) improves retrieval performance by enriching queries with related terms. Recently, Large Language Models (LLMs) have been used for QE, but existing methods face a trade-off: generating diverse terms boosts performance but increases computational cost. To address this challenge, we propose Candidate Token Query Expansion (CTQE), which extracts diverse and relevant terms from a single LLM decoding pass by leveraging unselected candidate tokens. These tokens, though not part of the final output, are conditioned on the full query and capture useful information. By aggregating them, CTQE achieves both relevance and diversity without extra inference, reducing overhead and latency. Experiments show that CTQE delivers strong retrieval performance with significantly lower cost, outperforming or comparable to more expensive methods. Code is available at: https://github.com/bluejeans8/CTQE

cs.IR↗

arXiv2025

Temporal Information Retrieval via Time-Specifier Model Merging

The rapid expansion of digital information and knowledge across structured and unstructured sources has heightened the importance of Information Retrieval (IR). While dense retrieval methods have substantially improved semantic matching for general queries, they consistently underperform on queries with explicit temporal constraints--often those containing numerical expressions and time specifiers such as ``in 2015.'' Existing approaches to Temporal Information Retrieval (TIR) improve temporal reasoning but often suffer from catastrophic forgetting, leading to reduced performance on non-temporal queries. To address this, we propose Time-Specifier Model Merging (TSM), a novel method that enhances temporal retrieval while preserving accuracy on non-temporal queries. TSM trains specialized retrievers for individual time specifiers and merges them in to a unified model, enabling precise handling of temporal constraints without compromising non-temporal retrieval. Extensive experiments on both temporal and non-temporal datasets demonstrate that TSM significantly improves performance on temporally constrained queries while maintaining strong results on non-temporal queries, consistently outperforming other baseline methods. Our code is available at https://github.com/seungyoonee/TSM .

cs.IR↗

arXiv2025

Does Rationale Quality Matter? Enhancing Mental Disorder Detection via Selective Reasoning Distillation

The detection of mental health problems from social media and the interpretation of these results have been extensively explored. Research has shown that incorporating clinical symptom information into a model enhances domain expertise, improving its detection and interpretation performance. While large language models (LLMs) are shown to be effective for generating explanatory rationales in mental health detection, their substantially large parameter size and high computational cost limit their practicality. Reasoning distillation transfers this ability to smaller language models (SLMs), but inconsistencies in the relevance and domain alignment of LLM-generated rationales pose a challenge. This paper investigates how rationale quality impacts SLM performance in mental health detection and explanation generation. We hypothesize that ensuring high-quality and domain-relevant rationales enhances the distillation. To this end, we propose a framework that selects rationales based on their alignment with expert clinical reasoning. Experiments show that our quality-focused approach significantly enhances SLM performance in both mental disorder detection and rationale generation. This work highlights the importance of rationale quality and offers an insightful framework for knowledge transfer in mental health applications.

cs.CL↗

arXiv2025

The RAG Paradox: A Black-Box Attack Exploiting Unintentional Vulnerabilities in Retrieval-Augmented Generation Systems

With the growing adoption of retrieval-augmented generation (RAG) systems, various attack methods have been proposed to degrade their performance. However, most existing approaches rely on unrealistic assumptions in which external attackers have access to internal components such as the retriever. To address this issue, we introduce a realistic black-box attack based on the RAG paradox, a structural vulnerability arising from the system's effort to enhance trust by revealing both the retrieved documents and their sources to users. This transparency enables attackers to observe which sources are used and how information is phrased, allowing them to craft poisoned documents that are more likely to be retrieved and upload them to the identified sources. Moreover, as RAG systems directly provide retrieved content to users, these documents must not only be retrievable but also appear natural and credible to maintain user confidence in the search results. Unlike prior work that focuses solely on improving document retrievability, our attack method explicitly considers both retrievability and user trust in the retrieved content. Both offline and online experiments demonstrate that our method significantly degrades system performance without internal access, while generating natural-looking poisoned documents.

cs.CR↗

arXiv2025

Lossless Acceleration of Large Language Models with Hierarchical Drafting based on Temporal Locality in Speculative Decoding

Accelerating inference in Large Language Models (LLMs) is critical for real-time interactions, as they have been widely incorporated into real-world services. Speculative decoding, a fully algorithmic solution, has gained attention for improving inference speed by drafting and verifying tokens, thereby generating multiple tokens in a single forward pass. However, current drafting strategies usually require significant fine-tuning or have inconsistent performance across tasks. To address these challenges, we propose Hierarchy Drafting (HD), a novel lossless drafting approach that organizes various token sources into multiple databases in a hierarchical framework based on temporal locality. In the drafting step, HD sequentially accesses multiple databases to obtain draft tokens from the highest to the lowest locality, ensuring consistent acceleration across diverse tasks and minimizing drafting latency. Our experiments on Spec-Bench using LLMs with 7B and 13B parameters demonstrate that HD outperforms existing database drafting methods, achieving robust inference speedups across model sizes, tasks, and temperatures.

cs.CL↗

arXiv2025

PiLaMIM: Toward Richer Visual Representations by Integrating Pixel and Latent Masked Image Modeling

In Masked Image Modeling (MIM), two primary methods exist: Pixel MIM and Latent MIM, each utilizing different reconstruction targets, raw pixels and latent representations, respectively. Pixel MIM tends to capture low-level visual details such as color and texture, while Latent MIM focuses on high-level semantics of an object. However, these distinct strengths of each method can lead to suboptimal performance in tasks that rely on a particular level of visual features. To address this limitation, we propose PiLaMIM, a unified framework that combines Pixel MIM and Latent MIM to integrate their complementary strengths. Our method uses a single encoder along with two distinct decoders: one for predicting pixel values and another for latent representations, ensuring the capture of both high-level and low-level visual features. We further integrate the CLS token into the reconstruction process to aggregate global context, enabling the model to capture more semantic information. Extensive experiments demonstrate that PiLaMIM outperforms key baselines such as MAE, I-JEPA and BootMAE in most cases, proving its effectiveness in extracting richer visual representations.

cs.CV↗

arXiv2024

EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation

We introduce EXIT, an extractive context compression framework that enhances both the effectiveness and efficiency of retrieval-augmented generation (RAG) in question answering (QA). Current RAG systems often struggle when retrieval models fail to rank the most relevant documents, leading to the inclusion of more context at the expense of latency and accuracy. While abstractive compression methods can drastically reduce token counts, their token-by-token generation process significantly increases end-to-end latency. Conversely, existing extractive methods reduce latency but rely on independent, non-adaptive sentence selection, failing to fully utilize contextual information. EXIT addresses these limitations by classifying sentences from retrieved documents - while preserving their contextual dependencies - enabling parallelizable, context-aware extraction that adapts to query complexity and retrieval quality. Our evaluations on both single-hop and multi-hop QA tasks show that EXIT consistently surpasses existing compression methods and even uncompressed baselines in QA accuracy, while also delivering substantial reductions in inference time and token count. By improving both effectiveness and efficiency, EXIT provides a promising direction for developing scalable, high-quality QA solutions in RAG pipelines. Our code is available at https://github.com/ThisIsHwang/EXIT

cs.CL↗

arXiv2024

Towards Effective Counter-Responses: Aligning Human Preferences with Strategies to Combat Online Trolling

Trolling in online communities typically involves disruptive behaviors such as provoking anger and manipulating discussions, leading to a polarized atmosphere and emotional distress. Robust moderation is essential for mitigating these negative impacts and maintaining a healthy and constructive community atmosphere. However, effectively addressing trolls is difficult because their behaviors vary widely and require different response strategies (RSs) to counter them. This diversity makes it challenging to choose an appropriate RS for each specific situation. To address this challenge, our research investigates whether humans have preferred strategies tailored to different types of trolling behaviors. Our findings reveal a correlation between the types of trolling encountered and the preferred RS. In this paper, we introduce a methodology for generating counter-responses to trolls by recommending appropriate RSs, supported by a dataset aligning these strategies with human preferences across various troll contexts. The experimental results demonstrate that our proposed approach guides constructive discussion and reduces the negative effects of trolls, thereby enhancing the online community environment.

cs.CL↗

arXiv2024

An Efficient Sign Language Translation Using Spatial Configuration and Motion Dynamics with LLMs

Gloss-free Sign Language Translation (SLT) converts sign videos directly into spoken language sentences without relying on glosses. Recently, Large Language Models (LLMs) have shown remarkable translation performance in gloss-free methods by harnessing their powerful natural language generation capabilities. However, these methods often rely on domain-specific fine-tuning of visual encoders to achieve optimal results. By contrast, this paper emphasizes the importance of capturing the spatial configurations and motion dynamics inherent in sign language. With this in mind, we introduce Spatial and Motion-based Sign Language Translation (SpaMo), a novel LLM-based SLT framework. The core idea of SpaMo is simple yet effective. We first extract spatial and motion features using off-the-shelf visual encoders and then input these features into an LLM with a language prompt. Additionally, we employ a visual-text alignment process as a warm-up before the SLT supervision. Our experiments demonstrate that SpaMo achieves state-of-the-art performance on two popular datasets, PHOENIX14T and How2Sign.

cs.CL↗

arXiv2024

DSLR: Document Refinement with Sentence-Level Re-ranking and Reconstruction to Enhance Retrieval-Augmented Generation

Recent advancements in Large Language Models (LLMs) have significantly improved their performance across various Natural Language Processing (NLP) tasks. However, LLMs still struggle with generating non-factual responses due to limitations in their parametric memory. Retrieval-Augmented Generation (RAG) systems address this issue by incorporating external knowledge with a retrieval module. Despite their successes, however, current RAG systems face challenges with retrieval failures and the limited ability of LLMs to filter out irrelevant information. Therefore, in this work, we propose DSLR (Document Refinement with Sentence-Level Re-ranking and Reconstruction), an unsupervised framework that decomposes retrieved documents into sentences, filters out irrelevant sentences, and reconstructs them again into coherent passages. We experimentally validate DSLR on multiple open-domain QA datasets and the results demonstrate that DSLR significantly enhances the RAG performance over conventional fixed-size passage. Furthermore, our DSLR enhances performance in specific, yet realistic scenarios without the need for additional training, providing an effective and efficient solution for refining retrieved documents in RAG systems.

cs.CL↗

arXiv2024

A Spatio-Temporal Representation Learning as an Alternative to Traditional Glosses in Sign Language Translation and Production

This work addresses the challenges associated with the use of glosses in both Sign Language Translation (SLT) and Sign Language Production (SLP). While glosses have long been used as a bridge between sign language and spoken language, they come with two major limitations that impede the advancement of sign language systems. First, annotating the glosses is a labor-intensive and time-consuming process, which limits the scalability of datasets. Second, the glosses oversimplify sign language by stripping away its spatio-temporal dynamics, reducing complex signs to basic labels and missing the subtle movements essential for precise interpretation. To address these limitations, we introduce Universal Gloss-level Representation (UniGloR), a framework designed to capture the spatio-temporal features inherent in sign language, providing a more dynamic and detailed alternative to the use of the glosses. The core idea of UniGloR is simple yet effective: We derive dense spatio-temporal representations from sign keypoint sequences using self-supervised learning and seamlessly integrate them into SLT and SLP tasks. Our experiments in a keypoint-based setting demonstrate that UniGloR either outperforms or matches the performance of previous SLT and SLP methods on two widely-used datasets: PHOENIX14T and How2Sign.

cs.CL↗

arXiv2024

Database-Augmented Query Representation for Information Retrieval

Information retrieval models that aim to search for documents relevant to a query have shown multiple successes, which have been applied to diverse tasks. Yet, the query from the user is oftentimes short, which challenges the retrievers to correctly fetch relevant documents. To tackle this, previous studies have proposed expanding the query with a couple of additional (user-related) features related to it. However, they may be suboptimal to effectively augment the query, and there is plenty of other information available to augment it in a relational database. Motivated by this fact, we present a novel retrieval framework called Database-Augmented Query representation (DAQu), which augments the original query with various (query-related) metadata across multiple tables. In addition, as the number of features in the metadata can be very large and there is no order among them, we encode them with the graph-based set-encoding strategy, which considers hierarchies of features in the database without order. We validate our DAQu in diverse retrieval scenarios, demonstrating that it significantly enhances overall retrieval performance over relevant baselines.

cs.CL↗

arXiv2024

Self-Knowledge Distillation for Learning Ambiguity

Recent language models have shown remarkable performance on natural language understanding (NLU) tasks. However, they are often sub-optimal when faced with ambiguous samples that can be interpreted in multiple ways, over-confidently predicting a single label without consideration for its correctness. To address this issue, we propose a novel self-knowledge distillation method that enables models to learn label distributions more accurately by leveraging knowledge distilled from their lower layers. This approach also includes a learning phase that re-calibrates the unnecessarily strengthened confidence for training samples judged as extremely ambiguous based on the distilled distribution knowledge. We validate our method on diverse NLU benchmark datasets and the experimental results demonstrate its effectiveness in producing better label distributions. Particularly, through the process of re-calibrating the confidence for highly ambiguous samples, the issue of over-confidence when predictions for unseen samples do not match with their ground-truth labels has been significantly alleviated. This has been shown to contribute to generating better distributions than the existing state-of-the-art method. Moreover, our method is more efficient in training the models compared to the existing method, as it does not involve additional training processes to refine label distributions.

cs.CL↗

arXiv2024

Typos that Broke the RAG's Back: Genetic Attack on RAG Pipeline by Simulating Documents in the Wild via Low-level Perturbations

The robustness of recent Large Language Models (LLMs) has become increasingly crucial as their applicability expands across various domains and real-world applications. Retrieval-Augmented Generation (RAG) is a promising solution for addressing the limitations of LLMs, yet existing studies on the robustness of RAG often overlook the interconnected relationships between RAG components or the potential threats prevalent in real-world databases, such as minor textual errors. In this work, we investigate two underexplored aspects when assessing the robustness of RAG: 1) vulnerability to noisy documents through low-level perturbations and 2) a holistic evaluation of RAG robustness. Furthermore, we introduce a novel attack method, the Genetic Attack on RAG (\textit{GARAG}), which targets these aspects. Specifically, GARAG is designed to reveal vulnerabilities within each component and test the overall system functionality against noisy documents. We validate RAG robustness by applying our \textit{GARAG} to standard QA datasets, incorporating diverse retrievers and LLMs. The experimental results show that GARAG consistently achieves high attack success rates. Also, it significantly devastates the performance of each component and their synergy, highlighting the substantial risk that minor textual inaccuracies pose in disrupting RAG systems in the real world.

cs.CL↗

arXiv2024

Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity

Retrieval-Augmented Large Language Models (LLMs), which incorporate the non-parametric knowledge from external knowledge bases into LLMs, have emerged as a promising approach to enhancing response accuracy in several tasks, such as Question-Answering (QA). However, even though there are various approaches dealing with queries of different complexities, they either handle simple queries with unnecessary computational overhead or fail to adequately address complex multi-step queries; yet, not all user requests fall into only one of the simple or complex categories. In this work, we propose a novel adaptive QA framework, that can dynamically select the most suitable strategy for (retrieval-augmented) LLMs from the simplest to the most sophisticated ones based on the query complexity. Also, this selection process is operationalized with a classifier, which is a smaller LM trained to predict the complexity level of incoming queries with automatically collected labels, obtained from actual predicted outcomes of models and inherent inductive biases in datasets. This approach offers a balanced strategy, seamlessly adapting between the iterative and single-step retrieval-augmented LLMs, as well as the no-retrieval methods, in response to a range of query complexities. We validate our model on a set of open-domain QA datasets, covering multiple query complexities, and show that ours enhances the overall efficiency and accuracy of QA systems, compared to relevant baselines including the adaptive retrieval approaches. Code is available at: https://github.com/starsuzi/Adaptive-RAG.

cs.CL↗