arXiv subjects

Jaeyoung Do

Publications and source records attributed to Jaeyoung Do.

33 records · Page 2Linked to original sources

arXiv2025

MatKV: Trading Compute for Flash Storage in LLM Inference

We observe two major trends in LLM-based generative AI: (1) inference is becoming the dominant factor in terms of cost and power consumption, surpassing training, and (2) retrieval augmented generation (RAG) is becoming prevalent. When processing long inputs in RAG, the prefill phase of computing the key-value vectors of input text is energy-intensive and time-consuming even with high-end GPUs. Thus, it is crucial to make the prefill phase in RAG inference efficient. To address this issue, we propose MatKV, a scheme that precomputes the key-value vectors (KVs) of RAG objects (e.g., documents), materializes them in inexpensive but fast and power-efficient flash storage, and reuses them at inference time instead of recomputing the KVs using costly and power-inefficient GPU. Experimental results using Hugging Face's Transformers library across state-of-the-art GPUs and flash memory SSDs confirm that, compared to full KV computation on GPUs, MatKV reduces both inference time and power consumption by half for RAG workloads, without severely impacting accuracy in the question-answering task. Furthermore, we demonstrate that MatKV enables additional optimizations in two ways. First, a GPU can decode text while simultaneously loading the materialized KVs for the next instance, reducing load latency. Second, since decoding speed is less sensitive to GPU performance than KV computation, low-end GPUs can be leveraged for decoding without significantly compromising speed once the materialized KVs are loaded into GPU memory. These findings underscore MatKV's potential to make large-scale generative AI applications more cost-effective, power-efficient, and accessible across a wider range of tasks and hardware environments.

cs.DC↗

arXiv2025

Don't Let It Fade: Preserving Edits in Diffusion Language Models via Token Timestep Allocation

While diffusion language models (DLMs) enable fine-grained refinement, their practical controllability remains fragile. We identify and formally characterize a central failure mode called update forgetting, in which uniform and context agnostic updates induce token level fluctuations across timesteps, erasing earlier semantic edits and disrupting the cumulative refinement process, thereby degrading fluency and coherence. As this failure originates in uniform and context agnostic updates, effective control demands explicit token ordering. We propose Token Timestep Allocation (TTA), which realizes soft and semantic token ordering via per token timestep schedules: critical tokens are frozen early, while uncertain tokens receive continued refinement. This timestep based ordering can be instantiated as either a fixed policy or an adaptive policy driven by task signals, thereby supporting a broad spectrum of refinement strategies. Because it operates purely at inference time, it applies uniformly across various DLMs and naturally extends to diverse supervision sources. Empirically, TTA improves controllability and fluency: on sentiment control, it yields more than 20 percent higher accuracy and nearly halves perplexity using less than one fifth the steps; in detoxification, it lowers maximum toxicity (12.2 versus 14.5) and perplexity (26.0 versus 32.0). Together, these results demonstrate that softened ordering via timestep allocation is the critical lever for mitigating update forgetting and achieving stable and controllable diffusion text generation.

cs.CL↗

arXiv2025

Exploring and Leveraging Class Vectors for Classifier Editing

Image classifiers play a critical role in detecting diseases in medical imaging and identifying anomalies in manufacturing processes. However, their predefined behaviors after extensive training make post hoc model editing difficult, especially when it comes to forgetting specific classes or adapting to distribution shifts. Existing classifier editing methods either focus narrowly on correcting errors or incur extensive retraining costs, creating a bottleneck for flexible editing. Moreover, such editing has seen limited investigation in image classification. To overcome these challenges, we introduce Class Vectors, which capture class-specific representation adjustments during fine-tuning. Whereas task vectors encode task-level changes in weight space, Class Vectors disentangle each class's adaptation in the latent space. We show that Class Vectors capture each class's semantic shift and that classifier editing can be achieved either by steering latent features along these vectors or by mapping them into weight space to update the decision boundaries. We also demonstrate that the inherent linearity and orthogonality of Class Vectors support efficient, flexible, and high-level concept editing via simple class arithmetic. Finally, we validate their utility in applications such as unlearning, environmental adaptation, adversarial defense, and adversarial trigger optimization.

cs.CV↗

arXiv2025

MMPB: It's Time for Multi-Modal Personalization

Visual personalization is essential in user-facing AI systems such as smart homes and healthcare, where aligning model behavior with user-centric concepts is critical. However, recent large Vision-Language Models (VLMs), despite their broad applicability, remain underexplored in their ability to adapt to individual users. In this paper, we introduce MMPB, the first extensive benchmark for evaluating VLMs on personalization. MMPB comprises 10k image-query pairs and includes 111 personalizable concepts across four categories: humans, animals, objects, and characters, with the human category enriched with preference-grounded queries. We structure personalization into three main task types, each highlighting a different key property of VLMs. Using 23 widely used VLMs including both open- and closed-source models, we evaluate personalization performance via a three-stage protocol: concept injection, multi-turn dialogue, and personalized querying. Our findings indicate that most VLMs (including some closed-source models) struggle with personalization, particularly in maintaining consistency over dialogue, handling user preferences, and adapting to visual cues. Our analysis reveals that the challenges in VLM personalization (such as refusal behaviors and long-context forgetting) highlight substantial room for improvement. By identifying these limitations and offering a scalable benchmark, MMPB offers valuable insights and a solid foundation for future research toward truly personalized multi-modal AI. Project Page: aidaslab.github.io/MMPB

cs.CV↗

arXiv2025

SECOND: Mitigating Perceptual Hallucination in Vision-Language Models via Selective and Contrastive Decoding

Despite significant advancements in Vision-Language Models (VLMs), the performance of existing VLMs remains hindered by object hallucination, a critical challenge to achieving accurate visual understanding. To address this issue, we propose SECOND: Selective and Contrastive Decoding, a novel approach that enables VLMs to effectively leverage multi-scale visual information with an object-centric manner, closely aligning with human visual perception. SECOND progressively selects and integrates multi-scale visual information, facilitating a more precise interpretation of images. By contrasting these visual information iteratively, SECOND significantly reduces perceptual hallucinations and outperforms a wide range of benchmarks. Our theoretical analysis and experiments highlight the largely unexplored potential of multi-scale application in VLMs, showing that prioritizing and contrasting across scales outperforms existing methods.

cs.CV↗

arXiv2025

MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula

In various academic and professional settings, such as mathematics lectures or research presentations, it is often necessary to convey mathematical expressions orally. However, reading mathematical expressions aloud without accompanying visuals can significantly hinder comprehension, especially for those who are hearing-impaired or rely on subtitles due to language barriers. For instance, when a presenter reads Euler's Formula, current Automatic Speech Recognition (ASR) models often produce a verbose and error-prone textual description (e.g., e to the power of i x equals cosine of x plus i $\textit{side}$ of x), instead of the concise $\LaTeX{}$ format (i.e., $ e^{ix} = \cos(x) + i\sin(x) $), which hampers clear understanding and communication. To address this issue, we introduce MathSpeech, a novel pipeline that integrates ASR models with small Language Models (sLMs) to correct errors in mathematical expressions and accurately convert spoken expressions into structured $\LaTeX{}$ representations. Evaluated on a new dataset derived from lecture recordings, MathSpeech demonstrates $\LaTeX{}$ generation capabilities comparable to leading commercial Large Language Models (LLMs), while leveraging fine-tuned small language models of only 120M parameters. Specifically, in terms of CER, BLEU, and ROUGE scores for $\LaTeX{}$ translation, MathSpeech demonstrated significantly superior capabilities compared to GPT-4o. We observed a decrease in CER from 0.390 to 0.298, and higher ROUGE/BLEU scores compared to GPT-4o.

cs.CL↗

arXiv2025

MathReader : Text-to-Speech for Mathematical Documents

TTS (Text-to-Speech) document reader from Microsoft, Adobe, Apple, and OpenAI have been serviced worldwide. They provide relatively good TTS results for general plain text, but sometimes skip contents or provide unsatisfactory results for mathematical expressions. This is because most modern academic papers are written in LaTeX, and when LaTeX formulas are compiled, they are rendered as distinctive text forms within the document. However, traditional TTS document readers output only the text as it is recognized, without considering the mathematical meaning of the formulas. To address this issue, we propose MathReader, which effectively integrates OCR, a fine-tuned T5 model, and TTS. MathReader demonstrated a lower Word Error Rate (WER) than existing TTS document readers, such as Microsoft Edge and Adobe Acrobat, when processing documents containing mathematical formulas. MathReader reduced the WER from 0.510 to 0.281 compared to Microsoft Edge, and from 0.617 to 0.281 compared to Adobe Acrobat. This will significantly contribute to alleviating the inconvenience faced by users who want to listen to documents, especially those who are visually impaired. The code is available at https://github.com/hyeonsieun/MathReader.

cs.AI↗

arXiv2024

MathBridge: A Large Corpus Dataset for Translating Spoken Mathematical Expressions into $LaTeX$ Formulas for Improved Readability

Improving the readability of mathematical expressions in text-based document such as subtitle of mathematical video, is an significant task. To achieve this, mathematical expressions should be convert to compiled formulas. For instance, the spoken expression ``x equals minus b plus or minus the square root of b squared minus four a c, all over two a'' from automatic speech recognition is more readily comprehensible when displayed as a compiled formula $x = \frac{-b \pm \sqrt{b^2 - 4ac}}{2a}$. To convert mathematical spoken sentences to compiled formulas, two processes are required: spoken sentences are converted into LaTeX formulas, and LaTeX formulas are converted into compiled formulas. The latter can be managed by using LaTeX engines. However, there is no way to do the former effectively. Even if we try to solve this using language models, there is no paired data between spoken sentences and LaTeX formulas to train it. In this paper, we introduce MathBridge, the first extensive dataset for translating mathematical spoken sentences into LaTeX formulas. MathBridge comprises approximately 23 million LaTeX formulas paired with the corresponding mathematical spoken sentences. Through comprehensive evaluations, including fine-tuning with proposed data, we discovered that MathBridge significantly enhances the capabilities of pretrained language models for converting to LaTeX formulas from mathematical spoken sentences. Specifically, for the T5-large model, the sacreBLEU score increased from 4.77 to 46.8, demonstrating substantial enhancement.

cs.LG↗

arXiv2024

Aligning Large Language Models via Fine-grained Supervision

Pre-trained large-scale language models (LLMs) excel at producing coherent articles, yet their outputs may be untruthful, toxic, or fail to align with user expectations. Current approaches focus on using reinforcement learning with human feedback (RLHF) to improve model alignment, which works by transforming coarse human preferences of LLM outputs into a feedback signal that guides the model learning process. However, because this approach operates on sequence-level feedback, it lacks the precision to identify the exact parts of the output affecting user preferences. To address this gap, we propose a method to enhance LLM alignment through fine-grained token-level supervision. Specifically, we ask annotators to minimally edit less preferred responses within the standard reward modeling dataset to make them more favorable, ensuring changes are made only where necessary while retaining most of the original content. The refined dataset is used to train a token-level reward model, which is then used for training our fine-grained Proximal Policy Optimization (PPO) model. Our experiment results demonstrate that this approach can achieve up to an absolute improvement of $5.1\%$ in LLM performance, in terms of win rate against the reference model, compared with the traditional PPO model.

cs.CL↗

arXiv2023

Data Augmentation for Improving Tail-traffic Robustness in Skill-routing for Dialogue Systems

Large-scale conversational systems typically rely on a skill-routing component to route a user request to an appropriate skill and interpretation to serve the request. In such system, the agent is responsible for serving thousands of skills and interpretations which create a long-tail distribution due to the natural frequency of requests. For example, the samples related to play music might be a thousand times more frequent than those asking for theatre show times. Moreover, inputs used for ML-based skill routing are often a heterogeneous mix of strings, embedding vectors, categorical and scalar features which makes employing augmentation-based long-tail learning approaches challenging. To improve the skill-routing robustness, we propose an augmentation of heterogeneous skill-routing data and training targeted for robust operation in long-tail data regimes. We explore a variety of conditional encoder-decoder generative frameworks to perturb original data fields and create synthetic training data. To demonstrate the effectiveness of the proposed method, we conduct extensive experiments using real-world data from a commercial conversational system. Based on the experiment results, the proposed approach improves more than 80% (51 out of 63) of intents with less than 10K of traffic instances in the skill-routing replication task.

cs.CL↗

arXiv2023

Scalable and Safe Remediation of Defective Actions in Self-Learning Conversational Systems

Off-Policy reinforcement learning has been a driving force for the state-of-the-art conversational AIs leading to more natural humanagent interactions and improving the user satisfaction for goal-oriented agents. However, in large-scale commercial settings, it is often challenging to balance between policy improvements and experience continuity on the broad spectrum of applications handled by such system. In the literature, off-policy evaluation and guard-railing on aggregate statistics has been commonly used to address this problem. In this paper, we propose a method for curating and leveraging high-precision samples sourced from historical regression incident reports to validate, safe-guard, and improve policies prior to the online deployment. We conducted extensive experiments using data from a real-world conversational system and actual regression incidents. The proposed method is currently deployed in our production system to protect customers against broken experiences and enable long-term policy improvements.

cs.AI↗

arXiv2022

Debiasing Neighbor Aggregation for Graph Neural Network in Recommender Systems

Graph neural networks (GNNs) have achieved remarkable success in recommender systems by representing users and items based on their historical interactions. However, little attention was paid to GNN's vulnerability to exposure bias: users are exposed to a limited number of items so that a system only learns a biased view of user preference to result in suboptimal recommendation quality. Although inverse propensity weighting is known to recognize and alleviate exposure bias, it usually works on the final objective with the model outputs, whereas GNN can also be biased during neighbor aggregation. In this paper, we propose a simple but effective approach, neighbor aggregation via inverse propensity (Navip) for GNNs. Specifically, given a user-item bipartite graph, we first derive propensity score of each user-item interaction in the graph. Then, inverse of the propensity score with Laplacian normalization is applied to debias neighbor aggregation from exposure bias. We validate the effectiveness of our approach through our extensive experiments on two public and Amazon Alexa datasets where the performance enhances up to 14.2%.

cs.IR↗

arXiv2022

Accelerating Large-Scale Graph-based Nearest Neighbor Search on a Computational Storage Platform

K-nearest neighbor search is one of the fundamental tasks in various applications and the hierarchical navigable small world (HNSW) has recently drawn attention in large-scale cloud services, as it easily scales up the database while offering fast search. On the other hand, a computational storage device (CSD) that combines programmable logic and storage modules on a single board becomes popular to address the data bandwidth bottleneck of modern computing systems. In this paper, we propose a computational storage platform that can accelerate a large-scale graph-based nearest neighbor search algorithm based on SmartSSD CSD. To this end, we modify the algorithm more amenable on the hardware and implement two types of accelerators using HLS- and RTL-based methodology with various optimization methods. In addition, we scale up the proposed platform to have 4 SmartSSDs and apply graph parallelism to boost the system performance further. As a result, the proposed computational storage platform achieves 75.59 query per second throughput for the SIFT1B dataset at 258.66W power dissipation, which is 12.83x and 17.91x faster and 10.43x and 24.33x more energy efficient than the conventional CPU-based and GPU-based server platform, respectively. With multi-terabyte storage and custom acceleration capability, we believe that the proposed computational storage platform is a promising solution for cost-sensitive cloud datacenters.

cs.AR↗

arXiv2020

ALEX: An Updatable Adaptive Learned Index

Recent work on "learned indexes" has changed the way we look at the decades-old field of DBMS indexing. The key idea is that indexes can be thought of as "models" that predict the position of a key in a dataset. Indexes can, thus, be learned. The original work by Kraska et al. shows that a learned index beats a B+Tree by a factor of up to three in search time and by an order of magnitude in memory footprint. However, it is limited to static, read-only workloads. In this paper, we present a new learned index called ALEX which addresses practical issues that arise when implementing learned indexes for workloads that contain a mix of point lookups, short range queries, inserts, updates, and deletes. ALEX effectively combines the core insights from learned indexes with proven storage and indexing techniques to achieve high performance and low memory footprint. On read-only workloads, ALEX beats the learned index from Kraska et al. by up to 2.2X on performance with up to 15X smaller index size. Across the spectrum of read-write workloads, ALEX beats B+Trees by up to 4.1X while never performing worse, with up to 2000X smaller index size. We believe ALEX presents a key step towards making learned indexes practical for a broader class of database workloads with dynamic updates.

cs.DB↗

arXiv2020

Lessons learned from the early performance evaluation of Intel Optane DC Persistent Memory in DBMS

Non-volatile memory (NVM) is an emerging technology, which has the persistence characteristics of large capacity storage devices(e.g., HDDs and SSDs), while providing the low access latency and byte-addressablity of traditional DRAM memory. This unique combination of features open up several new design considerations when building database management systems (DBMSs), such as replacing DRAM (as the main working space memory) or block devices (as the persistent storage), or complementing both at the same time for several DBMS components (such as access methods,storage engine, buffer management, logging/recovery, etc). However, interacting with NVM requires changes to application software to best use the device (e.g. mmap and clflush of small cache lines instead of write and fsync of large page buffers). Before introducing (potentially major) code changes to the DBMS for NVM, developers need a clear understanding of NVM performance in various conditions to help make better design choices. In this paper, we provide extensive performance evaluations conducted with a recently released NVM device, Intel Optane DC Persistent Memory (PMem), under different configurations with several micro-benchmark tools. Further, we evaluate OLTP and OLAP database workloads (i.e., TPC-C and TPC-H) with Microsoft SQL Server 2019 when using the NVM device as an in-memory buffer pool or persistent storage. From the lessons learned we share some recommendations for future DBMS design with PMem, e.g.simple hardware or software changes are not enough for the best use of PMem in DBMSs.

cs.DB↗