arXiv subjects

Youngsuk Park

Publications and source records attributed to Youngsuk Park.

At least 19 recordsLinked to original sources

arXiv2026

MKEvolve: A Modular Multi-Agent Framework for Kernel Code Generation

Despite rapid progress in LLM-based code generation, writing correct and performant kernels for hardware accelerators remains a key bottleneck in scaling modern ML workloads. We present MKEvolve (Modular Kernel Evolve), a framework that iteratively co-evolves a modular decomposition of complex PyTorch modules and the LLM-generated kernel for each submodule, refining the decomposition by splitting and fusing across iterations while independently improving each subkernel via LLM-driven beam search. The resulting kernels are programmatic compositions of independently verified subkernels, making them configurable (subkernel implementations are swappable), interpretable (errors and speedups are traceable to specific subkernels), and readily adaptable to related model architectures. Experiments with Triton on KernelBench L2 and L3, spanning multi-operator sequences and full model architectures, show that MKEvolve improves both correctness and speedup over end-to-end direct synthesis baselines while reducing LLM token usage by up to 35%.

cs.AI↗

arXiv2026

Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models

Reward hacking arises when a model improves a proxy reward by exploiting shortcuts rather than solving the intended task. We study this failure mode through the geometry of reinforcement learning updates in language models and argue that hacking emerges when optimization drifts away from a stable low-dimensional learning trajectory. We analyze this drift through dominant singular directions of parameter updates and show that reward-hacking runs exhibit substantially larger directional change than clean runs. Motivated by this observation, we introduce trusted-direction projection, which constrains gradients to remain within a clean reference subspace. Across reward-hacking experiments on mathematical reasoning, the proposed approach delays shortcut exploitation and better preserves task performance.

cs.LG↗

arXiv2026

Demystifying Transition Matching: When and Why It Can Beat Flow Matching

Flow Matching (FM) underpins many state-of-the-art generative models, yet recent results indicate that Transition Matching (TM) can achieve higher quality with fewer sampling steps. This work answers the question of when and why TM outperforms FM. First, when the target is a unimodal Gaussian distribution, we prove that TM attains strictly lower KL divergence than FM for finite number of steps. The improvement arises from stochastic difference latent updates in TM, which preserve target covariance that deterministic FM underestimates. We then characterize convergence rates, showing that TM achieves faster convergence than FM under a fixed compute budget, establishing its advantage in the unimodal Gaussian setting. Second, we extend the analysis to Gaussian mixtures and identify local-unimodality regimes in which the sampling dynamics approximate the unimodal case, where TM can outperform FM. The approximation error decreases as the minimal distance between component means increases, highlighting that TM is favored when the modes are well separated. However, when the target variance approaches zero, each TM update converges to the FM update, and the performance advantage of TM diminishes. In summary, we show that TM outperforms FM when the target distribution has well-separated modes and non-negligible variances. We validate our theoretical results with controlled experiments on Gaussian distributions, and extend the comparison to real-world applications in image and video generation.

cs.LG↗

arXiv2026

Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs

Scaling the number of parameters and the size of training data has proven to be an effective strategy for improving large language model (LLM) performance. Yet, as these models grow increasingly powerful and widely deployed, the cost of inference has become a pressing concern. Despite its importance, the trade-off between model accuracy and inference efficiency remains underexplored. In this work, we examine how key architectural factors, hidden size, the allocation of parameters between MLP and attention (mlp-to-attention ratio), and grouped-query attention (GQA), influence both inference cost and accuracy. We introduce a conditional scaling law that augments the Chinchilla framework with architectural information, along with a search framework for identifying architectures that are simultaneously inference-efficient and accurate. To validate our approach, we train more than 200 models spanning 80M to 3B parameters and 8B to 100B training tokens, and fit the proposed conditional scaling law. Our results show that the conditional scaling law reliably predicts optimal architectural choices and that the resulting models outperform existing open-source baselines. Under the same training budget, optimized architectures achieve up to 2.1% higher accuracy and 42% greater inference throughput compared to LLaMA-3.2.

cs.LG↗

arXiv2026

Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMs

Speculative decoding is widely used in accelerating large language model (LLM) inference. In this work, we focus on the online draft model selection problem in speculative decoding. We design an algorithm that provably competes with the best draft model in hindsight for each query in terms of either the token acceptance probability or expected acceptance length. In particular, we show that we can accurately evaluate all draft models, instead of only the chosen model without incurring additional queries to the target model, which allows us to improve exponentially over the existing bandit-based approach as the number of draft models increases. Our approach is generically applicable with any speculative decoding methods (single draft, multi-drafts and draft-trees). Moreover, we design system-efficient versions of online learners and demonstrate that the overhead in computation and latency can be substantially reduced. We conduct extensive experiments on open-source LLMs and diverse datasets, demonstrating that our methods substantially outperform the state-of-the-art EAGLE3 and the BanditSpec baseline in a variety of domains where specialized domain-expert drafters are available, especially when long reasoning chains are required.

cs.LG↗

arXiv2026

Online Posterior Sampling with a Diffusion Prior

Posterior sampling in contextual bandits with a Gaussian prior can be implemented exactly or approximately using the Laplace approximation. The Gaussian prior is computationally efficient but it cannot describe complex distributions. In this work, we propose approximate posterior sampling algorithms for contextual bandits with a diffusion model prior. The key idea is to sample from a chain of approximate conditional posteriors, one for each stage of the reverse diffusion process, which are obtained by the Laplace approximation. Our approximations are motivated by posterior sampling with a Gaussian prior, and inherit its simplicity and efficiency. They are asymptotically consistent and perform well empirically on a variety of contextual bandit problems.

cs.LG↗

arXiv2026

TritonRL: Training LLMs to Think and Code Triton Without Cheating

The rapid evolution of Large Language Models (LLMs) has driven a growing demand for automated, high-performance system kernels to accelerate machine learning workloads. We introduce TritonRL, a domain-specialized 8B-scale LLM for Triton programming, trained via a novel reinforcement learning (RL) framework. While Triton synthesis faces unique challenges, including data scarcity and a high susceptibility to reward hacking, our approach enables robust kernel generation through two primary innovations. First, we implement a multi-layered verification system that provides high-fidelity reward signals, ensuring that generated kernels are both syntactically and functionally valid. Second, we propose Hierarchical Reward Decomposition (HRD), which decouples reinforcement for high-level reasoning and low-level implementation to resolve the credit assignment problem in long-sequence generation. Comprehensive evaluations on KernelBench demonstrate that TritonRL achieves state-of-the-art correctness and runtime speedup, outperforming concurrent Triton-specific models and matching the performance of frontier models with over 100B parameters. Our results highlight the effectiveness of hardware-aware RL paradigms in specialized domain adaptation.

cs.SE↗

arXiv2025

Theoretical Guarantees of Learning Ensembling Strategies with Applications to Time Series Forecasting

Ensembling is among the most popular tools in machine learning (ML) due to its effectiveness in minimizing variance and thus improving generalization. Most ensembling methods for black-box base learners fall under the umbrella of "stacked generalization," namely training an ML algorithm that takes the inferences from the base learners as input. While stacking has been widely applied in practice, its theoretical properties are poorly understood. In this paper, we prove a novel result, showing that choosing the best stacked generalization from a (finite or finite-dimensional) family of stacked generalizations based on cross-validated performance does not perform "much worse" than the oracle best. Our result strengthens and significantly extends the results in Van der Laan et al. (2007). Inspired by the theoretical analysis, we further propose a particular family of stacked generalizations in the context of probabilistic forecasting, each one with a different sensitivity for how much the ensemble weights are allowed to vary across items, timestamps in the forecast horizon, and quantiles. Experimental results demonstrate the performance gain of the proposed method.

cs.LG↗

arXiv2025

MuonBP: Faster Muon via Block-Periodic Orthogonalization

Gradient orthogonalization is a simple strategy that shows great utility in speeding up gradient descent. The Muon optimizer (Jordan, Jin, et al., 2024) combines gradient orthogonalization with first-order momentum and achieves significant improvement in data efficiency over Adam/AdamW (Loshchilov and Hutter, 2019) for language model training. However, when using model parallelism, gradient orthogonalization introduces additional overhead compared to coordinate-wise optimizers (such as AdamW) due to additional gather and scatter operations on gradient matrix shards from different devices. This additional communication can amount to a throughput hit of 5%-10% compared to Adam/AdamW. To remedy this, we propose Muon with Block-Periodic Orthogonalization (MuonBP), which applies orthogonalization independently to matrix shards on each device and periodically performs full orthogonalization to maintain training stability at scale. We show how to adjust the learning rate from the baseline to MuonBP and give convergence guarantees for this algorithm. Crucially, our theory dictates that we use two stepsizes: one for the blockwise orthogonalization steps, and one for the full orthogonalization steps. Our method is simple, requires minimal hyperparameter adjustments, and achieves competitive iteration complexity compared with baseline Muon while providing per-iteration throughput comparable to coordinate-wise methods such as AdamW. When training an 8B model with eight-way tensor parallelism and ZeRO optimizer state sharding, MuonBP achieves 8% throughput increase compared to Muon with no degradation in performance.

cs.LG↗

arXiv2025

Training LLMs with MXFP4

Low precision (LP) datatypes such as MXFP4 can accelerate matrix multiplications (GEMMs) and reduce training costs. However, directly using MXFP4 instead of BF16 during training significantly degrades model quality. In this work, we present the first near-lossless training recipe that uses MXFP4 GEMMs, which are $2\times$ faster than FP8 on supported hardware. Our key insight is to compute unbiased gradient estimates with stochastic rounding (SR), resulting in more accurate model updates. However, directly applying SR to MXFP4 can result in high variance from block-level outliers, harming convergence. To overcome this, we use the random Hadamard tranform to theoretically bound the variance of SR. We train GPT models up to 6.7B parameters and find that our method induces minimal degradation over mixed-precision BF16 training. Our recipe computes $>1/2$ the training FLOPs in MXFP4, enabling an estimated speedup of $>1.3\times$ over FP8 and $>1.7\times$ over BF16 during backpropagation.

cs.LG↗

arXiv2025

ProxSparse: Regularized Learning of Semi-Structured Sparsity Masks for Pretrained LLMs

Large Language Models (LLMs) have demonstrated exceptional performance in natural language processing tasks, yet their massive size makes serving them inefficient and costly. Semi-structured pruning has emerged as an effective method for model acceleration, but existing approaches are suboptimal because they focus on local, layer-wise optimizations using heuristic rules, failing to leverage global feedback. We present ProxSparse, a learning-based framework for mask selection enabled by regularized optimization. ProxSparse transforms the rigid, non-differentiable mask selection process into a smoother optimization procedure, allowing gradual mask exploration with flexibility. ProxSparse does not involve additional weight updates once the mask is determined. Our extensive evaluations on 7 widely used models show that ProxSparse consistently outperforms previously proposed semi-structured mask selection methods with significant improvement, demonstrating the effectiveness of our learned approach towards semi-structured pruning.

cs.LG↗

arXiv2025

RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models

Supervised fine-tuning is a standard method for adapting pre-trained large language models (LLMs) to downstream tasks. Quantization has been recently studied as a post-training technique for efficient LLM deployment. To obtain quantized fine-tuned LLMs, conventional pipelines would first fine-tune the pre-trained models, followed by post-training quantization. This often yields suboptimal performance as it fails to leverage the synergy between fine-tuning and quantization. To effectively realize low-bit quantization of weights, activations and KV caches in LLMs, we propose an algorithm named Rotated Straight-Through-Estimator (RoSTE), which combines quantization-aware supervised fine-tuning (QA-SFT) with an adaptive rotation strategy that identifies an effective rotation configuration to reduce activation outliers. We provide theoretical insights on RoSTE by analyzing its prediction error when applied to an overparameterized least square quantized training problem. Our findings reveal that the prediction error is directly proportional to the quantization error of the converged weights, which can be effectively managed through an optimized rotation configuration. Experiments on Pythia, Qwen and Llama models of different sizes demonstrate the effectiveness of RoSTE. Compared to existing post-SFT quantization baselines, our method consistently achieves superior performances across various tasks and different LLM architectures. Our code is available at https://github.com/OptimAI-Lab/RoSTE.

cs.LG↗

arXiv2025

Stochastic Rounding for LLM Training: Theory and Practice

As the parameters of Large Language Models (LLMs) have scaled to hundreds of billions, the demand for efficient training methods -- balancing faster computation and reduced memory usage without sacrificing accuracy -- has become more critical than ever. In recent years, various mixed precision strategies, which involve different precision levels for optimization components, have been proposed to increase training speed with minimal accuracy degradation. However, these strategies often require manual adjustments and lack theoretical justification. In this work, we leverage stochastic rounding (SR) to address numerical errors of training with low-precision representation. We provide theoretical analyses of implicit regularization and convergence under the Adam optimizer when SR is utilized. With the insights from these analyses, we extend previous BF16 + SR strategy to be used in distributed settings, enhancing the stability and performance for large scale training. Empirical results from pre-training models with up to 6.7B parameters, for the first time, demonstrate that our BF16 with SR strategy outperforms (BF16, FP32) mixed precision strategies, achieving better validation perplexity, up to $1.54\times$ higher throughput, and $30\%$ less memory usage.

cs.LG↗

arXiv2024

Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization

How to best develop foundational models for time series forecasting remains an important open question. Tokenization is a crucial consideration in this effort: what is an effective discrete vocabulary for a real-valued sequential input? To address this question, we develop WaveToken, a wavelet-based tokenizer that allows models to learn complex representations directly in the space of time-localized frequencies. Our method first scales and decomposes the input time series, then thresholds and quantizes the wavelet coefficients, and finally pre-trains an autoregressive model to forecast coefficients for the forecast horizon. By decomposing coarse and fine structures in the inputs, wavelets provide an eloquent and compact language for time series forecasting that simplifies learning. Empirical results on a comprehensive benchmark, including 42 datasets for both in-domain and zero-shot settings, show that WaveToken: i) provides better accuracy than recently proposed foundation models for forecasting while using a much smaller vocabulary (1024 tokens), and performs on par or better than modern deep learning models trained specifically on each dataset; and ii) exhibits superior generalization capabilities, achieving the best average rank across all datasets for three complementary metrics. In addition, we show that our method can easily capture complex temporal patterns of practical relevance that are challenging for other recent pre-trained models, including trends, sparse spikes, and non-stationary time series with varying frequencies evolving over time.

cs.LG↗

arXiv2024

Inference Optimization of Foundation Models on AI Accelerators

Powerful foundation models, including large language models (LLMs), with Transformer architectures have ushered in a new era of Generative AI across various industries. Industry and research community have witnessed a large number of new applications, based on those foundation models. Such applications include question and answer, customer services, image and video generation, and code completions, among others. However, as the number of model parameters reaches to hundreds of billions, their deployment incurs prohibitive inference costs and high latency in real-world scenarios. As a result, the demand for cost-effective and fast inference using AI accelerators is ever more higher. To this end, our tutorial offers a comprehensive discussion on complementary inference optimization techniques using AI accelerators. Beginning with an overview of basic Transformer architectures and deep learning system frameworks, we deep dive into system optimization techniques for fast and memory-efficient attention computations and discuss how they can be implemented efficiently on AI accelerators. Next, we describe architectural elements that are key for fast transformer inference. Finally, we examine various model compression and fast decoding strategies in the same context.

cs.AI↗

arXiv2024

Collage: Light-Weight Low-Precision Strategy for LLM Training

Large models training is plagued by the intense compute cost and limited hardware memory. A practical solution is low-precision representation but is troubled by loss in numerical accuracy and unstable training rendering the model less useful. We argue that low-precision floating points can perform well provided the error is properly compensated at the critical locations in the training process. We propose Collage which utilizes multi-component float representation in low-precision to accurately perform operations with numerical errors accounted. To understand the impact of imprecision to training, we propose a simple and novel metric which tracks the lost information during training as well as differentiates various precision strategies. Our method works with commonly used low-precision such as half-precision ($16$-bit floating points) and can be naturally extended to work with even lower precision such as $8$-bit. Experimental results show that pre-training using Collage removes the requirement of using $32$-bit floating-point copies of the model and attains similar/better training performance compared to $(16, 32)$-bit mixed-precision strategy, with up to $3.7\times$ speedup and $\sim 15\%$ to $23\%$ less memory usage in practice.

cs.LG↗

arXiv2024

Variance-reduced Zeroth-Order Methods for Fine-Tuning Language Models

Fine-tuning language models (LMs) has demonstrated success in a wide array of downstream tasks. However, as LMs are scaled up, the memory requirements for backpropagation become prohibitively high. Zeroth-order (ZO) optimization methods can leverage memory-efficient forward passes to estimate gradients. More recently, MeZO, an adaptation of ZO-SGD, has been shown to consistently outperform zero-shot and in-context learning when combined with suitable task prompts. In this work, we couple ZO methods with variance reduction techniques to enhance stability and convergence for inference-based LM fine-tuning. We introduce Memory-Efficient Zeroth-Order Stochastic Variance-Reduced Gradient (MeZO-SVRG) and demonstrate its efficacy across multiple LM fine-tuning tasks, eliminating the reliance on task-specific prompts. Evaluated across a range of both masked and autoregressive LMs on benchmark GLUE tasks, MeZO-SVRG outperforms MeZO with up to 20% increase in test accuracies in both full- and partial-parameter fine-tuning settings. MeZO-SVRG benefits from reduced computation time as it often surpasses MeZO's peak test accuracy with a $2\times$ reduction in GPU-hours. MeZO-SVRG significantly reduces the required memory footprint compared to first-order SGD, i.e. by $2\times$ for autoregressive models. Our experiments highlight that MeZO-SVRG's memory savings progressively improve compared to SGD with larger batch sizes.

cs.LG↗

arXiv2023

Robust Multivariate Time-Series Forecasting: Adversarial Attacks and Defense Mechanisms

This work studies the threats of adversarial attack on multivariate probabilistic forecasting models and viable defense mechanisms. Our studies discover a new attack pattern that negatively impact the forecasting of a target time series via making strategic, sparse (imperceptible) modifications to the past observations of a small number of other time series. To mitigate the impact of such attack, we have developed two defense strategies. First, we extend a previously developed randomized smoothing technique in classification to multivariate forecasting scenarios. Second, we develop an adversarial training algorithm that learns to create adversarial examples and at the same time optimizes the forecasting model to improve its robustness against such adversarial simulation. Extensive experiments on real-world datasets confirm that our attack schemes are powerful and our defense algorithms are more effective compared with baseline defense mechanisms.

cs.LG↗