arXiv subjects

Eunhyeok Park

Publications and source records attributed to Eunhyeok Park.

At least 19 recordsLinked to original sources

arXiv2026

LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Models

Looped Language Models (LoopLMs) perform "latent reasoning" by recursively refining internal latent representations with shared weights, offering a more effective alternative to explicit verbal reasoning. Despite their effectiveness, we find that LoopLMs remain prone to loop instability: unstable refinement across iterations can produce localized uncertain "hard" tokens associated with reasoning errors. To address this, we propose LoopCD, loop-wise contrastive decoding that enhances the reasoning performance of LoopLMs by intervening on these tokens at inference time. Specifically, we exploit the internal dynamics of LoopLMs and contrast the logits from earlier iterations with logits from the last refined iteration to form the final sampling distribution. We find that this strategy is highly efficient, introducing only negligible inference overhead and requiring no additional training, while effectively improving reasoning performance by naturally refining reasoning-critical hard tokens. Extensive experiments show that our method improves the performance of recent representative LoopLMs across various reasoning tasks.

cs.CL↗

arXiv2026

TaRA: Training-Aware Low-Rank Adaptation Initialization

Low-Rank Adaptation (LoRA) has become a de facto standard for parameter-efficient fine-tuning (PEFT), yet its performance is highly sensitive to initialization due to the information bottleneck imposed by low-rank decomposition. Existing approaches attempt to construct high-quality LoRA initializations by exploiting principal components of pretrained weights, activations, or gradients. However, these methods do not directly account for the training dynamics of the full-rank model. In this paper, we propose Training-aware Low-Rank Adaptation Initialization (TaRA), a method that initializes LoRA such that the gradients induced by the low-rank factors closely approximate the gradient of the corresponding full-rank weight matrix. Derived from a mathematical formulation, TaRA improves gradient fidelity at the start of training while introducing negligible computational overhead. Across diverse and challenging fine-tuning tasks, TaRA consistently outperforms prior state-of-the-art methods, establishing a simple, robust, and scalable solution for effective LoRA initialization.

cs.CL↗

arXiv2026

AdaHOP: Fast and Accurate Low-Precision Training via Outlier-Pattern-Aware Rotation

Hadamard transforms have become a key tool for stabilizing low-precision training, but existing methods apply them uniformly across tensors and computation paths. We show that this one-size-fits-all strategy is inherently limited: Hadamard smoothing reduces quantization error only when its direction is properly aligned with the operand's outlier structure. Through a systematic study of weights, activations, and gradients in LLM training, we identify three stable outlier patterns, Row-wise, Column-wise, and None, and show that each outlier pattern pair in matrix multiplication requires a distinct transform or outlier-handling strategy. We propose AdaHOP, Adaptive Hadamard transform with Outlier-Pattern-aware strategy, which applies Inner Hadamard Transform (IHT) when inner-dimension mixing properly suppresses the operands' outliers, and selectively applies Outlier Extraction (OE) that extracts dominant outlier rows or columns into a high-precision path when it does not. With fused, hardware-aware Triton kernels, AdaHOP enables training from scratch at MXFP4 precision with BF16-level quality, while achieving up to 3.6X memory compression, 1.46X end-to-end training speedup over BF16.

cs.LG↗

arXiv2026

Speculative Coupled Decoding for Training-Free Lossless Acceleration of Autoregressive Visual Generation

Autoregressive (AR) modeling has recently emerged as a promising new paradigm in visual generation, but its practical adoption is severely constrained by the slow inference speed of per-token generation, which often requires thousands of steps to produce a single sample. While several Speculative Decoding (SD)-based methods have been proposed to solve this problem by generating multiple tokens in a single forward step, they suffer from limited speedup, degraded quality, or require the training of a draft model. To solve these problems, we propose a new training-free, lossless SD framework, Speculative Coupled Decoding (SCD), by extending the recently proposed Speculative Jacobi Decoding (SJD). While SJD shows strong potential for accelerating AR generation by combining Jacobi iteration and SD, we found that its acceptance rate is still significantly limited due to the instability arising from the independent sampling process used during draft token generation. To overcome this, we introduce an information-theoretic approach, Coupling, which stabilizes the drafting trajectory of SJD by maximizing the probability of sampling identical draft tokens across consecutive iterations, significantly enhancing the acceptance rate while preserving its lossless property. Remarkably, this method requires only a single-line modification to the existing algorithm with almost zero overhead, yet achieves substantial performance gains, delivering up to a 4.2x speedup in image generation and 13.6x speedup in video generation compared to standard AR decoding, without any degradation or the need for additional training. The source code is available at https://github.com/junhyukso/SCD

cs.CV↗

arXiv2026

GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning

Low-Rank Adaptation (LoRA) is a popular method for parameter-efficient fine-tuning (PEFT) of generative models, valued for its simplicity and effectiveness. Despite recent enhancements, LoRA still suffers from a fundamental limitation: overfitting when the bottleneck is widened. It performs best at ranks 32-64, yet its accuracy stagnates or declines at higher ranks, still falling short of full fine-tuning (FFT) performance. We identify the root cause as LoRA's structural bottleneck, which introduces gradient entanglement to the unrelated input channels and distorts gradient propagation. To address this, we introduce a novel structure, Granular Low-Rank Adaptation (GraLoRA) that partitions weight matrices into sub-blocks, each with its own low-rank adapter. With negligible computational or storage cost, GraLoRA overcomes LoRA's limitations, effectively increases the representational capacity, and more closely approximates FFT behavior. Experiments on code generation and commonsense reasoning benchmarks show that GraLoRA consistently outperforms LoRA and other baselines, achieving up to +8.5% absolute gain in Pass@1 on HumanEval+. These improvements hold across model sizes and rank settings, making GraLoRA a scalable and robust solution for PEFT. Code, data, and scripts are available at https://github.com/SqueezeBits/GraLoRA.git

cs.LG↗

arXiv2025

Stabilizing Direct Training of Spiking Neural Networks: Membrane Potential Initialization and Threshold-robust Surrogate Gradient

Recent advancements in the direct training of Spiking Neural Networks (SNNs) have demonstrated high-quality outputs even at early timesteps, paving the way for novel energy-efficient AI paradigms. However, the inherent non-linearity and temporal dependencies in SNNs introduce persistent challenges, such as temporal covariate shift (TCS) and unstable gradient flow with learnable neuron thresholds. In this paper, we present two key innovations: MP-Init (Membrane Potential Initialization) and TrSG (Threshold-robust Surrogate Gradient). MP-Init addresses TCS by aligning the initial membrane potential with its stationary distribution, while TrSG stabilizes gradient flow with respect to threshold voltage during training. Extensive experiments validate our approach, achieving state-of-the-art accuracy on both static and dynamic image datasets. The code is available at: https://github.com/kookhh0827/SNN-MP-Init-TRSG

cs.NE↗

arXiv2025

Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking

Generative Behavior Cloning (GBC) is a simple yet effective framework for robot learning, particularly in multi-task settings. Recent GBC methods often employ diffusion policies with open-loop (OL) control, where actions are generated via a diffusion process and executed in multi-step chunks without replanning. While this approach has demonstrated strong success rates and generalization, its inherent stochasticity can result in erroneous action sampling, occasionally leading to unexpected task failures. Moreover, OL control suffers from delayed responses, which can degrade performance in noisy or dynamic environments. To address these limitations, we propose two novel techniques to enhance the consistency and reactivity of diffusion policies: (1) self-guidance, which improves action fidelity by leveraging past observations and implicitly promoting future-aware behavior; and (2) adaptive chunking, which selectively updates action sequences when the benefits of reactivity outweigh the need for temporal consistency. Extensive experiments show that our approach substantially improves GBC performance across a wide range of simulated and real-world robotic manipulation tasks. Our code is available at https://github.com/junhyukso/SGAC

cs.RO↗

arXiv2025

PruneCD: Contrasting Pruned Self Model to Improve Decoding Factuality

To mitigate the hallucination problem in large language models, DoLa exploits early exit logits from the same model as a contrastive prior. However, we found that these early exit logits tend to be flat, low in magnitude, and fail to reflect meaningful contrasts. To address this, we propose PruneCD, a novel contrastive decoding method that constructs the amateur model via layer pruning rather than early exit. This design leads to more informative and well-aligned logits, enabling more effective contrastive decoding. Through qualitative and quantitative analyses, we demonstrate that PruneCD consistently improves factuality with minimal inference overhead, offering a robust and practical approach to mitigating hallucinations in LLMs.

cs.CL↗

arXiv2025

AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models

To enable broader deployment of Large Language Models (LLMs), it is essential to identify the best-performing model under strict memory constraints. We present AMQ, Automated Mixed-Precision Weight-Only Quantization, a framework that assigns layer-wise quantization bit-widths to optimally balance model quality and memory usage. However, the combinatorial search space, with over 10^{100} possible configurations, makes conventional black-box optimization infeasible. AMQ overcomes this challenge through four key innovations:(1) search space pruning using prior knowledge to exclude unpromising configurations, (2) quantization proxy to bypass costly format conversions during search, (3) quality predictor to minimize evaluation overhead, and (4) iterative search-and-update strategy for fast and stable convergence. By integrating these components, AMQ efficiently explores the quality-efficiency landscape, reaching the Pareto frontier and yielding LLMs that are both compact and high-performing. Our code is available at https://github.com/dlwns147/amq.

cs.LG↗

arXiv2025

Déjà Vu: Efficient Video-Language Query Engine with Learning-based Inter-Frame Computation Reuse

Recently, Video-Language Models (VideoLMs) have demonstrated remarkable capabilities, offering significant potential for flexible and powerful video query systems. These models typically rely on Vision Transformers (ViTs), which process video frames individually to extract visual embeddings. However, generating embeddings for large-scale videos requires ViT inferencing across numerous frames, posing a major hurdle to real-world deployment and necessitating solutions for integration into scalable video data management systems. This paper introduces Déjà Vu, a video-language query engine that accelerates ViT-based VideoLMs by reusing computations across consecutive frames. At its core is ReuseViT, a modified ViT model specifically designed for VideoLM tasks, which learns to detect inter-frame reuse opportunities, striking an effective balance between accuracy and reuse. Although ReuseViT significantly reduces computation, these savings do not directly translate into performance gains on GPUs. To overcome this, Déjà Vu integrates memory-compute joint compaction techniques that convert the FLOP savings into tangible performance gains. Evaluations on three VideoLM tasks show that Déjà Vu accelerates embedding generation by up to a 2.64x within a 2% error bound, dramatically enhancing the practicality of VideoLMs for large-scale video analytics.

cs.DC↗

arXiv2025

Grouped Speculative Decoding for Autoregressive Image Generation

Recently, autoregressive (AR) image models have demonstrated remarkable generative capabilities, positioning themselves as a compelling alternative to diffusion models. However, their sequential nature leads to long inference times, limiting their practical scalability. In this work, we introduce Grouped Speculative Decoding (GSD), a novel, training-free acceleration method for AR image models. While recent studies have explored Speculative Decoding (SD) as a means to speed up AR image generation, existing approaches either provide only modest acceleration or require additional training. Our in-depth analysis reveals a fundamental difference between language and image tokens: image tokens exhibit inherent redundancy and diversity, meaning multiple tokens can convey valid semantics. However, traditional SD methods are designed to accept only a single most-likely token, which fails to leverage this difference, leading to excessive false-negative rejections. To address this, we propose a new SD strategy that evaluates clusters of visually valid tokens rather than relying on a single target token. Additionally, we observe that static clustering based on embedding distance is ineffective, which motivates our dynamic GSD approach. Extensive experiments show that GSD accelerates AR image models by an average of 3.7x while preserving image quality-all without requiring any additional training. The source code is available at https://github.com/junhyukso/GSD

cs.CV↗

arXiv2025

SEAL: Scaling to Emphasize Attention for Long-Context Retrieval

While many advanced LLMs are designed to handle long sequence data, we can still observe notable quality degradation even within the sequence limit. In this work, we introduce a novel approach called Scaling to Emphasize Attention for Long-context retrieval (SEAL), which enhances the retrieval performance of large language models (LLMs) over long contexts. We observe that specific attention heads are closely tied to long-context retrieval, showing positive or negative correlation with retrieval scores, and adjusting the strength of these heads boosts the quality of LLMs in long context by a large margin. Built on this insight, we propose a learning-based mechanism that leverages generated data to emphasize these heads. By applying SEAL, we achieve significant improvements in long-context retrieval performance across various tasks and models. Additionally, when combined with existing training-free context extension techniques, SEAL extends the contextual limits of LLMs while maintaining highly reliable outputs.

cs.CL↗

arXiv2025

The Oversmoothing Fallacy: A Misguided Narrative in GNN Research

Oversmoothing has been recognized as a main obstacle to building deep Graph Neural Networks (GNNs), limiting the performance. This position paper argues that the influence of oversmoothing has been overstated and advocates for a further exploration of deep GNN architectures. Given the three core operations of GNNs, aggregation, linear transformation, and non-linear activation, we show that prior studies have mistakenly confused oversmoothing with the vanishing gradient, caused by transformation and activation rather than aggregation. Our finding challenges prior beliefs about oversmoothing being unique to GNNs. Furthermore, we demonstrate that classical solutions such as skip connections and normalization enable the successful stacking of deep GNN layers without performance degradation. Our results clarify misconceptions about oversmoothing and shed new light on the potential of deep GNNs.

cs.LG↗

arXiv2025

Efficient Latent Semantic Clustering for Scaling Test-Time Computation of LLMs

Scaling test-time computation--generating and analyzing multiple or sequential outputs for a single input--has become a promising strategy for improving the reliability and quality of large language models (LLMs), as evidenced by advances in uncertainty quantification and multi-step reasoning. A key shared component is semantic clustering, which groups outputs that differ in form but convey the same meaning. Semantic clustering enables estimation of the distribution over the semantics of outputs and helps avoid redundant exploration of reasoning paths. However, existing approaches typically rely on external models, which introduce substantial computational overhead and often fail to capture context-aware semantics. We propose Latent Semantic Clustering (LSC), a lightweight and context-sensitive method that leverages the generator LLM's internal hidden states for clustering, eliminating the need for external models. Our extensive experiment across various LLMs and datasets shows that LSC significantly improves the computational efficiency of test-time scaling while maintaining or exceeding the performance of existing methods.

cs.CL↗

arXiv2025

Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation

Model merging has emerged as a powerful technique for combining task-specific weights, achieving superior performance in multi-target domain adaptation. However, when applied to practical scenarios, such as quantized models, new challenges arise. In practical scenarios, quantization is often applied to target-specific data, but this process restricts the domain of interest and introduces discretization effects, making model merging highly non-trivial. In this study, we analyze the impact of quantization on model merging through the lens of error barriers. Leveraging these insights, we propose a novel post-training quantization, HDRQ - Hessian and distant regularizing quantization - that is designed to consider model merging for multi-target domain adaptation. Our approach ensures that the quantization process incurs minimal deviation from the source pre-trained model while flattening the loss surface to facilitate smooth model merging. To our knowledge, this is the first study on this challenge, and extensive experiments confirm its effectiveness.

cs.LG↗

arXiv2025

PTQ4VM: Post-Training Quantization for Visual Mamba

Visual Mamba is an approach that extends the selective space state model, Mamba, to vision tasks. It processes image tokens sequentially in a fixed order, accumulating information to generate outputs. Despite its growing popularity for delivering high-quality outputs at a low computational cost across various tasks, Visual Mamba is highly susceptible to quantization, which makes further performance improvements challenging. Our analysis reveals that the fixed token access order in Visual Mamba introduces unique quantization challenges, which we categorize into three main issues: 1) token-wise variance, 2) channel-wise outliers, and 3) a long tail of activations. To address these challenges, we propose Post-Training Quantization for Visual Mamba (PTQ4VM), which introduces two key strategies: Per-Token Static (PTS) quantization and Joint Learning of Smoothing Scale and Step Size (JLSS). To the our best knowledge, this is the first quantization study on Visual Mamba. PTQ4VM can be applied to various Visual Mamba backbones, converting the pretrained model to a quantized format in under 15 minutes without notable quality degradation. Extensive experiments on large-scale classification and regression tasks demonstrate its effectiveness, achieving up to 1.83x speedup on GPUs with negligible accuracy loss compared to FP16. Our code is available at https://github.com/YoungHyun197/ptq4vm.

cs.CV↗

arXiv2025

HOT: Hadamard-based Optimized Training

It has become increasingly important to optimize backpropagation to reduce memory usage and computational overhead. Achieving this goal is highly challenging, as multiple objectives must be considered jointly while maintaining training quality. In this paper, we focus on matrix multiplication, which accounts for the largest portion of training costs, and analyze its backpropagation in detail to identify lightweight techniques that offer the best benefits. Based on this analysis, we introduce a novel method, Hadamard-based Optimized Training (HOT). In this approach, we apply Hadamard-based optimizations, such as Hadamard quantization and Hadamard low-rank approximation, selectively and with awareness of the suitability of each optimization for different backward paths. Additionally, we introduce two enhancements: activation buffer compression and layer-wise quantizer selection. Our extensive analysis shows that HOT achieves up to 75% memory savings and a 2.6 times acceleration on real GPUs, with negligible accuracy loss compared to FP32 precision.

cs.LG↗

arXiv2025

PCM : Picard Consistency Model for Fast Parallel Sampling of Diffusion Models

Recently, diffusion models have achieved significant advances in vision, text, and robotics. However, they still face slow generation speeds due to sequential denoising processes. To address this, a parallel sampling method based on Picard iteration was introduced, effectively reducing sequential steps while ensuring exact convergence to the original output. Nonetheless, Picard iteration does not guarantee faster convergence, which can still result in slow generation in practice. In this work, we propose a new parallelization scheme, the Picard Consistency Model (PCM), which significantly reduces the number of generation steps in Picard iteration. Inspired by the consistency model, PCM is directly trained to predict the fixed-point solution, or the final output, at any stage of the convergence trajectory. Additionally, we introduce a new concept called model switching, which addresses PCM's limitations and ensures exact convergence. Extensive experiments demonstrate that PCM achieves up to a 2.71x speedup over sequential sampling and a 1.77x speedup over Picard iteration across various tasks, including image generation and robotic control.

cs.CV↗