Donghyun Lee

Mo Tiwari, Ryan Kang, Je-Yong Lee et al. · harvard

Maximum Inner Product Search (MIPS) is a ubiquitous task in machine learning applications such as recommendation systems. Given a query vector and $n$ atom vectors in $d$-dimensional space, the goal of MIPS is to find the atom that has the highest inner product with the query vector. Existing MIPS algorithms scale at least as $O(\sqrt{d})$, which becomes computationally prohibitive in high-dimensional settings. In this work, we present BanditMIPS, a novel randomized MIPS algorithm whose complexity is independent of $d$. BanditMIPS estimates the inner product for each atom by subsampling coordinates and adaptively evaluates more coordinates for more promising atoms. The specific adaptive sampling strategy is motivated by multi-armed bandits. We provide theoretical guarantees that BanditMIPS returns the correct answer with high probability, while improving the complexity in $d$ from $O(\sqrt{d})$ to $O(1)$. We also perform experiments on four synthetic and real-world datasets and demonstrate that BanditMIPS outperforms prior state-of-the-art algorithms. For example, in the Movie Lens dataset ($n$=4,000, $d$=6,000), BanditMIPS is 20$\times$ faster than the next best algorithm while returning the same answer. BanditMIPS requires no preprocessing of the data and includes a hyperparameter that practitioners may use to trade off accuracy and runtime. We also propose a variant of our algorithm, named BanditMIPS-$α$, which achieves further speedups by employing non-uniform sampling across coordinates. Finally, we demonstrate how known preprocessing techniques can be used to further accelerate BanditMIPS, and discuss applications to Matching Pursuit and Fourier analysis.

Lakshya A Agrawal, Donghyun Lee, Shangyin Tan et al.

Can a single LLM-based optimization system match specialized tools across fundamentally different domains? We show that when optimization problems are formulated as improving a text artifact evaluated by a scoring function, a single AI-based optimization system-supporting single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs-achieves state-of-the-art results across six diverse tasks. Our system discovers agent architectures that nearly triple Gemini Flash's ARC-AGI accuracy (32.5% to 89.5%), finds scheduling algorithms that cut cloud costs by 40%, generates CUDA kernels where 87% match or beat PyTorch, and outperforms AlphaEvolve's reported circle packing solution (n=26). Ablations across three domains reveal that actionable side information yields faster convergence and substantially higher final scores than score-only feedback, and that multi-task search outperforms independent optimization given equivalent per-problem budget through cross-task transfer, with benefits scaling with the number of related tasks. Together, we show for the first time that text optimization with LLM-based search is a general-purpose problem-solving paradigm, unifying tasks traditionally requiring domain-specific algorithms under a single framework. We open-source optimize\_anything with support for multiple backends as part of the GEPA project at https://github.com/gepa-ai/gepa .

Shiting Xiao, Yuhang Li, Youngeun Kim et al.

Spiking Neural Networks (SNNs) have emerged as a compelling, energy-efficient alternative to traditional Artificial Neural Networks (ANNs) for static image tasks such as image classification and segmentation. However, in the more complex video classification domain, SNN-based methods fall considerably short of ANN-based benchmarks due to the challenges in processing dense frame sequences. To bridge this gap, we propose ReSpike, a hybrid framework that synergizes the strengths of ANNs and SNNs to tackle action recognition tasks with high accuracy and low energy cost. By decomposing film clips into spatial and temporal components, i.e., RGB image Key Frames and event-like Residual Frames, ReSpike leverages ANN for learning spatial information and SNN for learning temporal information. In addition, we propose a multi-scale cross-attention mechanism for effective feature fusion. Compared to state-of-the-art SNN baselines, our ReSpike hybrid architecture demonstrates significant performance improvements (e.g., >30% absolute accuracy improvement on HMDB-51, UCF-101, and Kinetics-400). Furthermore, ReSpike achieves comparable performance with prior ANN approaches while bringing better accuracy-energy tradeoff.

Mo Tiwari, Ryan Kang, Donghyun Lee et al.

Clustering is a fundamental task in data science with wide-ranging applications. In $k$-medoids clustering, cluster centers must be actual datapoints and arbitrary distance metrics may be used; these features allow for greater interpretability of the cluster centers and the clustering of exotic objects in $k$-medoids clustering, respectively. $k$-medoids clustering has recently grown in popularity due to the discovery of more efficient $k$-medoids algorithms. In particular, recent research has proposed BanditPAM, a randomized $k$-medoids algorithm with state-of-the-art complexity and clustering accuracy. In this paper, we present BanditPAM++, which accelerates BanditPAM via two algorithmic improvements, and is $O(k)$ faster than BanditPAM in complexity and substantially faster than BanditPAM in wall-clock runtime. First, we demonstrate that BanditPAM has a special structure that allows the reuse of clustering information $\textit{within}$ each iteration. Second, we demonstrate that BanditPAM has additional structure that permits the reuse of information $\textit{across}$ different iterations. These observations inspire our proposed algorithm, BanditPAM++, which returns the same clustering solutions as BanditPAM but often several times faster. For example, on the CIFAR10 dataset, BanditPAM++ returns the same results as BanditPAM but runs over 10$\times$ faster. Finally, we provide a high-performance C++ implementation of BanditPAM++, callable from Python and R, that may be of interest to practitioners at https://github.com/motiwari/BanditPAM. Auxiliary code to reproduce all of our experiments via a one-line script is available at https://github.com/ThrunGroup/BanditPAM_plusplus_experiments.

Geonhee Ahn, Donghyun Lee, Hayoung Doo et al.

Large language models (LLMs) have enabled agentic AI systems for scientific discovery, but most approaches remain limited to textbased reasoning without automated experimental verification. We propose MIND, an LLM-driven framework for automated hypothesis validation in materials research. MIND organizes the scientific discovery process into hypothesis refinement, experimentation, and debate-based validation within a multi-agent pipeline. For experimental verification, the system integrates Machine Learning Interatomic Potentials, particularly SevenNet-Omni, enabling scalable in-silico experiments. We also provide a web-based user interface for automated hypothesis testing. The modular design allows additional experimental modules to be integrated, making the framework adaptable to broader scientific workflows. The code is available at: https://github.com/IMMS-Ewha/MIND, and a demonstration video at: https://youtu.be/lqiFe1OQzN4.

Yuhang Li, Ruokai Yin, Donghyun Lee et al.

We introduce GPTAQ, a novel finetuning-free quantization method for compressing large-scale transformer architectures. Unlike the previous GPTQ method, which independently calibrates each layer, we always match the quantized layer's output to the exact output in the full-precision model, resulting in a scheme that we call asymmetric calibration. Such a scheme can effectively reduce the quantization error accumulated in previous layers. We analyze this problem using optimal brain compression to derive a close-formed solution. The new solution explicitly minimizes the quantization error as well as the accumulated asymmetry error. Furthermore, we utilize various techniques to parallelize the solution calculation, including channel parallelization, neuron decomposition, and Cholesky reformulation for matrix fusion. As a result, GPTAQ is easy to implement, simply using 20 more lines of code than GPTQ but improving its performance under low-bit quantization. Remarkably, on a single GPU, we quantize a 405B language transformer as well as EVA-02, the rank first vision transformer that achieves 90% pretraining Imagenet accuracy. Code is available at Github.

Yuhang Li, Youngeun Kim, Donghyun Lee et al.

In the realm of deep neural network deployment, low-bit quantization presents a promising avenue for enhancing computational efficiency. However, it often hinges on the availability of training data to mitigate quantization errors, a significant challenge when data availability is scarce or restricted due to privacy or copyright concerns. Addressing this, we introduce GenQ, a novel approach employing an advanced Generative AI model to generate photorealistic, high-resolution synthetic data, overcoming the limitations of traditional methods that struggle to accurately mimic complex objects in extensive datasets like ImageNet. Our methodology is underscored by two robust filtering mechanisms designed to ensure the synthetic data closely aligns with the intrinsic characteristics of the actual training data. In case of limited data availability, the actual data is used to guide the synthetic data generation process, enhancing fidelity through the inversion of learnable token embeddings. Through rigorous experimentation, GenQ establishes new benchmarks in data-free and data-scarce quantization, significantly outperforming existing methods in accuracy and efficiency, thereby setting a new standard for quantization in low data regimes. Code is released at \url{https://github.com/Intelligent-Computing-Lab-Yale/GenQ}.

Ruokai Yin, Yuhang Li, Donghyun Lee et al.

Large language models (LLMs) deliver strong performance but are difficult to deploy due to high memory and compute costs. While pruning reduces these demands, most methods ignore activation sparsity observed at runtime. We reinterpret activation sparsity as dynamic structured weight sparsity and propose DuoGPT, a unified framework that constructs dual-sparse (spMspV) workloads by combining unstructured weight pruning with activation sparsity. To preserve accuracy, we extend the Optimal Brain Compression (OBC) framework with activation-aware calibration and introduce output residuals from the dense model as correction terms. We further optimize the solution for efficient GPU execution, enabling scalability to billion-parameter LLMs. Evaluations on LLaMA-2 and LLaMA-3 show that DuoGPT outperforms state-of-the-art structured pruning methods by up to 9.17% accuracy at an iso-speedup of 1.39$\times$ compared to the baseline dense model. Code is available at Github.

Hankyul Kang, Gregor Seifer, Donghyun Lee et al.

According to the forgetting curve theory, we can enhance memory retention by learning extensive data and taking adequate rest. This means that in order to effectively retain new knowledge, it is essential to learn it thoroughly and ensure sufficient rest so that our brain can memorize without forgetting. The main takeaway from this theory is that learning extensive data at once necessitates sufficient rest before learning the same data again. This aspect of human long-term memory retention can be effectively utilized to address the continual learning of neural networks. Retaining new knowledge for a long period of time without catastrophic forgetting is the critical problem of continual learning. Therefore, based on Ebbinghaus' theory, we introduce the view-batch model that adjusts the learning schedules to optimize the recall interval between retraining the same samples. The proposed view-batch model allows the network to get enough rest to learn extensive knowledge from the same samples with a recall interval of sufficient length. To this end, we specifically present two approaches: 1) a replay method that guarantees the optimal recall interval, and 2) a self-supervised learning that acquires extensive knowledge from a single training sample at a time. We empirically show that these approaches of our method are aligned with the forgetting curve theory, which can enhance long-term memory. In our experiments, we also demonstrate that our method significantly improves many state-of-the-art continual learning methods in various protocols and scenarios. We open-source this project at https://github.com/hankyul2/ViewBatchModel.

Shiting Xiao, Rishabh Kabra, Yuhang Li et al.

The ability to segment objects based on open-ended language prompts remains a critical challenge, requiring models to ground textual semantics into precise spatial masks while handling diverse and unseen categories. We present OpenWorldSAM, a framework that extends the prompt-driven Segment Anything Model v2 (SAM2) to open-vocabulary scenarios by integrating multi-modal embeddings extracted from a lightweight vision-language model (VLM). Our approach is guided by four key principles: i) Unified prompting: OpenWorldSAM supports a diverse range of prompts, including category-level and sentence-level language descriptions, providing a flexible interface for various segmentation tasks. ii) Efficiency: By freezing the pre-trained components of SAM2 and the VLM, we train only 4.5 million parameters on the COCO-stuff dataset, achieving remarkable resource efficiency. iii) Instance Awareness: We enhance the model's spatial understanding through novel positional tie-breaker embeddings and cross-attention layers, enabling effective segmentation of multiple instances. iv) Generalization: OpenWorldSAM exhibits strong zero-shot capabilities, generalizing well on unseen categories and an open vocabulary of concepts without additional training. Extensive experiments demonstrate that OpenWorldSAM achieves state-of-the-art performance in open-vocabulary semantic, instance, and panoptic segmentation across multiple benchmarks. Code is available at https://github.com/GinnyXiao/OpenWorldSAM.

Donghyun Lee, Yuhang Li, Ruokai Yin et al.

State Space Models (SSMs) have emerged as powerful alternatives to attention-based Transformers, with Mamba demonstrating impressive efficiency and scalability. As these models grow increasingly larger, the need for Parameter-Efficient Fine-Tuning (PEFT) methods becomes critical to adapt pre-trained Mamba to downstream tasks without prohibitive computational costs. However, previous approaches simply apply traditional Transformer-tailored PEFT methods without addressing the unique temporal processing dynamics of SSMs. To address this limitation, we propose Memba, a membrane-driven PEFT approach specifically designed for Mamba. Memba introduces Leaky Integrate Membrane (LIM) neurons as bio-inspired gating mechanisms that naturally accumulate membrane potentials over time, enhancing selective information retention. By strategically combining LIM neurons with Low-Rank Adaptations (LoRA) and cross-layer membrane transfer, our approach significantly improves Mamba's temporal modeling capabilities. Extensive experiments across language and vision tasks demonstrate that Memba achieves substantial improvements over existing PEFT methods. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/Memba.

Donghyun Lee, Je-Yong Lee, Genghan Zhang et al.

Large Language Models (LLMs) have dramatically advanced AI applications, yet their deployment remains challenging due to their immense inference costs. Recent studies ameliorate the computational costs of LLMs by increasing their activation sparsity but suffer from significant performance degradation on downstream tasks. In this work, we introduce a new framework for sparsifying the activations of base LLMs and reducing inference costs, dubbed Contextually Aware Thresholding for Sparsity (CATS). CATS is relatively simple, easy to implement, and highly effective. At the heart of our framework is a new non-linear activation function. We demonstrate that CATS can be applied to various base models, including Mistral-7B and Llama2-7B, and outperforms existing sparsification techniques in downstream task performance. More precisely, CATS-based models often achieve downstream task performance within 1-2% of their base models without any fine-tuning and even at activation sparsity levels of 50%. Furthermore, CATS-based models converge faster and display better task performance than competing techniques when fine-tuning is applied. Finally, we develop a custom GPU kernel for efficient implementation of CATS that translates the activation of sparsity of CATS to real wall-clock time speedups. Our custom kernel implementation of CATS results in a ~15% improvement in wall-clock inference latency of token generation on both Llama-7B and Mistral-7B.

Zeyi Huang, Yuyang Ji, Xiaofang Wang et al.

Long-form video understanding with Large Vision Language Models is challenged by the need to analyze temporally dispersed yet spatially concentrated key moments within limited context windows. In this work, we introduce VideoMindPalace, a new framework inspired by the "Mind Palace", which organizes critical video moments into a topologically structured semantic graph. VideoMindPalace organizes key information through (i) hand-object tracking and interaction, (ii) clustered activity zones representing specific areas of recurring activities, and (iii) environment layout mapping, allowing natural language parsing by LLMs to provide grounded insights on spatio-temporal and 3D context. In addition, we propose the Video MindPalace Benchmark (VMB), to assess human-like reasoning, including spatial localization, temporal reasoning, and layout-aware sequential understanding. Evaluated on VMB and established video QA datasets, including EgoSchema, NExT-QA, IntentQA, and the Active Memories Benchmark, VideoMindPalace demonstrates notable gains in spatio-temporal coherence and human-aligned reasoning, advancing long-form video analysis capabilities in VLMs.

Yuhang Li, Donghyun Lee, Ruokai Yin et al.

Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40\% better results than previous state-of-the-art decomposition methods, the SVD-LLM.

Donghyun Lee, Dawoon Jeong, Jae W. Lee et al.

The advent of 3D Gaussian Splatting has revolutionized graphics rendering by delivering high visual quality and fast rendering speeds. However, training large-scale scenes at high quality remains challenging due to the substantial memory demands required to store parameters, gradients, and optimizer states, which can quickly overwhelm GPU memory. To address these limitations, we propose GS-Scale, a fast and memory-efficient training system for 3D Gaussian Splatting. GS-Scale stores all Gaussians in host memory, transferring only a subset to the GPU on demand for each forward and backward pass. While this dramatically reduces GPU memory usage, it requires frustum culling and optimizer updates to be executed on the CPU, introducing slowdowns due to CPU's limited compute and memory bandwidth. To mitigate this, GS-Scale employs three system-level optimizations: (1) selective offloading of geometric parameters for fast frustum culling, (2) parameter forwarding to pipeline CPU optimizer updates with GPU computation, and (3) deferred optimizer update to minimize unnecessary memory accesses for Gaussians with zero gradients. Our extensive evaluations on large-scale datasets demonstrate that GS-Scale significantly lowers GPU memory demands by 3.3-5.6x, while achieving training speeds comparable to GPU without host offloading. This enables large-scale 3D Gaussian Splatting training on consumer-grade GPUs; for instance, GS-Scale can scale the number of Gaussians from 4 million to 18 million on an RTX 4070 Mobile GPU, leading to 23-35% LPIPS (learned perceptual image patch similarity) improvement.

Nak-Jun Sung, Donghyun Lee, Bo Hwa Choi et al.

Mammography report generation is a critical yet underexplored task in medical AI, characterized by challenges such as multiview image reasoning, high-resolution visual cues, and unstructured radiologic language. In this work, we introduce AMRG (Automatic Mammography Report Generation), the first end-to-end framework for generating narrative mammography reports using large vision-language models (VLMs). Building upon MedGemma-4B-it-a domain-specialized, instruction-tuned VLM-we employ a parameter-efficient fine-tuning (PEFT) strategy via Low-Rank Adaptation (LoRA), enabling lightweight adaptation with minimal computational overhead. We train and evaluate AMRG on DMID, a publicly available dataset of paired high-resolution mammograms and diagnostic reports. This work establishes the first reproducible benchmark for mammography report generation, addressing a longstanding gap in multimodal clinical AI. We systematically explore LoRA hyperparameter configurations and conduct comparative experiments across multiple VLM backbones, including both domain-specific and general-purpose models under a unified tuning protocol. Our framework demonstrates strong performance across both language generation and clinical metrics, achieving a ROUGE-L score of 0.5691, METEOR of 0.6152, CIDEr of 0.5818, and BI-RADS accuracy of 0.5582. Qualitative analysis further highlights improved diagnostic consistency and reduced hallucinations. AMRG offers a scalable and adaptable foundation for radiology report generation and paves the way for future research in multimodal medical AI.

Donghyun Lee, Dawoon Jeong, Jae W. Lee et al.

Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve 2.55x end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy.

Seonglae Cho, Harryn Oh, Donghyun Lee et al.

Sparse Autoencoders (SAEs) have emerged as a promising solution for decomposing large language model representations into interpretable features. However, Paulo and Belrose (2025) have highlighted instability across different initialization seeds, and Heap et al. (2025) have pointed out that SAEs may not capture model-internal features. These problems likely stem from training SAEs on external datasets - either collected from the Web or generated by another model - which may contain out-of-distribution (OOD) data beyond the model's generalisation capabilities. This can result in hallucinated SAE features, which we term "Fake Features", that misrepresent the model's internal activations. To address these issues, we propose FaithfulSAE, a method that trains SAEs on the model's own synthetic dataset. Using FaithfulSAEs, we demonstrate that training SAEs on less-OOD instruction datasets results in SAEs being more stable across seeds. Notably, FaithfulSAEs outperform SAEs trained on web-based datasets in the SAE probing task and exhibit a lower Fake Feature Ratio in 5 out of 7 models. Overall, our approach eliminates the dependency on external datasets, advancing interpretability by better capturing model-internal features while highlighting the often neglected importance of SAE training datasets.

Zeyi Huang, Utkarsh Ojha, Yuyang Ji et al.

When a human undertakes a test, their responses likely follow a pattern: if they answered an easy question $(2 \times 3)$ incorrectly, they would likely answer a more difficult one $(2 \times 3 \times 4)$ incorrectly; and if they answered a difficult question correctly, they would likely answer the easy one correctly. Anything else hints at memorization. Do current visual recognition models exhibit a similarly structured learning capacity? In this work, we consider the task of image classification and study if those models' responses follow that pattern. Since real images aren't labeled with difficulty, we first create a dataset of 100 categories, 10 attributes, and 3 difficulty levels using recent generative models: for each category (e.g., dog) and attribute (e.g., occlusion), we generate images of increasing difficulty (e.g., a dog without occlusion, a dog only partly visible). We find that most of the models do in fact behave similarly to the aforementioned pattern around 80-90% of the time. Using this property, we then explore a new way to evaluate those models. Instead of testing the model on every possible test image, we create an adaptive test akin to GRE, in which the model's performance on the current round of images determines the test images in the next round. This allows the model to skip over questions too easy/hard for itself, and helps us get its overall performance in fewer steps.

Yujee Song, Donghyun Lee, Rui Meng et al.

A Marked Temporal Point Process (MTPP) is a stochastic process whose realization is a set of event-time data. MTPP is often used to understand complex dynamics of asynchronous temporal events such as money transaction, social media, healthcare, etc. Recent studies have utilized deep neural networks to capture complex temporal dependencies of events and generate embedding that aptly represent the observed events. While most previous studies focus on the inter-event dependencies and their representations, how individual events influence the overall dynamics over time has been under-explored. In this regime, we propose a Decoupled MTPP framework that disentangles characterization of a stochastic process into a set of evolving influences from different events. Our approach employs Neural Ordinary Differential Equations (Neural ODEs) to learn flexible continuous dynamics of these influences while simultaneously addressing multiple inference problems, such as density estimation and survival rate computation. We emphasize the significance of disentangling the influences by comparing our framework with state-of-the-art methods on real-life datasets, and provide analysis on the model behavior for potential applications.

Donghyun Lee, Minkyoung Cho, Seungwon Lee et al.

Post-training quantization is a representative technique for compressing neural networks, making them smaller and more efficient for deployment on edge devices. However, an inaccessible user dataset often makes it difficult to ensure the quality of the quantized neural network in practice. In addition, existing approaches may use a single uniform bit-width across the network, resulting in significant accuracy degradation at extremely low bit-widths. To utilize multiple bit-width, sensitivity metric plays a key role in balancing accuracy and compression. In this paper, we propose a novel sensitivity metric that considers the effect of quantization error on task loss and interaction with other layers. Moreover, we develop labeled data generation methods that are not dependent on a specific operation of the neural network. Our experiments show that the proposed metric better represents quantization sensitivity, and generated data are more feasible to be applied to mixed-precision quantization.

Hosung Park, Donghyun Lee, Minkyu Lim et al.

In this paper, a time delay neural network (TDNN) based acoustic model is proposed to implement a fast-converged acoustic modeling for Korean speech recognition. The TDNN has an advantage in fast-convergence where the amount of training data is limited, due to subsampling which excludes duplicated weights. The TDNN showed an absolute improvement of 2.12% in terms of character error rate compared to feed forward neural network (FFNN) based modelling for Korean speech corpora. The proposed model converged 1.67 times faster than a FFNN-based model did.