arXiv subjects

Sunwoo Lee

Publications and source records attributed to Sunwoo Lee.

At least 19 recordsLinked to original sources

arXiv2026

A.X K2 Technical Report

We introduce A.X K2, a 688B-parameter Mixture-of-Experts (MoE) language model trained from scratch as a high-performance foundation for \emph{agentic} applications. Trained on approximately 8.5T tokens---fewer than its predecessor, A.X K1---on a smaller but higher-quality mixture with substantially expanded agentic and software-engineering data, it nonetheless improves over A.X K1 across the board, by over 30 percentage points on some benchmarks, reflecting large gains in token efficiency. To support long contexts efficiently, we introduce Sparse Gated Attention (SGA), which combines sparse attention with gated attention, and adopt Gated Norm (GN) to stabilize large-scale training. SGA is trained natively at 128K through a \emph{sparse} indexer warmup that optimizes the indexer against its own sparse top-$k$ selection rather than the dense attention distribution, making adaptation markedly cheaper: each query reads only 2,048 positions, yet long-context quality is unchanged and A.X K2 scores 94.6 on RULER out to 256K. The outlier suppression of GN in turn keeps 4-bit NVFP4 serving within one point of FP8 accuracy. A simple yet effective Think-Fusion recipe further lets users switch between thinking and non-thinking modes within a single unified model. Extensive evaluations show that A.X K2 performs competitively against strong open-weight baselines, matching or exceeding them on math and Korean-language benchmarks.

cs.AI↗

arXiv2026

Output-Aware Rotation for INT2 KV-Cache Quantization

The key-value (KV) cache has become a major memory and bandwidth bottleneck in long-context large language model inference, making ultra-low-bit quantization increasingly important. However, existing rotation-based INT2 methods optimize cache statistics or proxy errors before the complete attention readout, even though the model is ultimately affected by the error propagated through attention and the output projection $W_O$. To address this mismatch, we propose \textit{OptR}, an output-aware rotation method that minimizes post-$W_O$ attention-output error. OptR decomposes the post-$W_O$ attention-output error into key- and value-induced terms and learns per-head orthogonal corrections through the full INT2 quantization and attention path. OptR further applies an attention-equivalent key reparameterization to reduce large channel-wise offsets without changing the softmax distribution. Across three models and five reasoning and coding benchmarks, OptR consistently improves both QuaRot and OSCAR and strengthens long-context retrieval, while preserving the paged KV-cache format with negligible inference overhead.

cs.LG↗

arXiv2026

Locality-Aware Redundancy Pruning for LLM Depth Compression

Large language models are known to contain representational redundancy across network depth, making depth pruning an effective approach for improving inference efficiency. Existing one-shot pruning methods rely on local layer importance or fixed redundancy assumptions across architectures. We propose Locality-Aware Redundancy Pruning (LoRP), a training-free one-shot depth pruning framework guided by representation locality. We show that inter-layer redundancy can be either localized or globally distributed depending on the LLM architecture. To characterize this phenomenon, we introduce Representation Locality Score (RLS), derived from global inter-layer hidden-state similarity. Using a small calibration set, LoRP computes pairwise layer similarity, clusters layers by representational similarity, and allocates pruning according to residual intra-cluster redundancy. Experiments across diverse LLM families show improvements in both perplexity and downstream task accuracy. Official github repository: https://github.com/daniel-eai/LoRP-Locality-Aware-Redundancy-Pruning/

cs.LG↗

arXiv2026

Interaction-Breaking Adversarial Learning Framework for Robust Multi-Agent Reinforcement Learning

Cooperation is central to multi-agent reinforcement learning (MARL), yet learned coordination can be fragile when external perturbations disrupt inter-agent interactions. Prior robust MARL methods have primarily considered value-oriented attacks, leaving a gap in robustness when interaction structures themselves are corrupted. In this paper, we propose an interaction-breaking adversarial learning (IBAL) framework that takes an information-theoretic view to construct attacks that impede coordination by perturbing agents' observations and actions, and trains agents to perform reliably under such disruptions. Empirically, our approach improves robustness over existing robust MARL baselines across diverse attack settings and yields stronger performance even under agent-missing scenarios. Our code is available at https://sunwoolee0504.github.io/IBAL.

cs.LG↗

arXiv2026

Ghosted Layers: Unconstrained Activation Alignment for Recovering Layer-Pruned LLMs

Layer pruning removes entire Transformer decoder blocks from large language models, but introduces a mismatch between the hidden state received by the next surviving layer and the distribution it was trained to process, leading to significant performance degradation. We propose Ghosted Layers, a training-free recovery module that addresses this issue by solving a boundary activation alignment problem. Our method derives a closed-form optimal linear operator from a small calibration set to reconstruct the activation discrepancy introduced by the pruned layers. We show that this solution corresponds to the unconstrained optimum of the alignment objective, whereas existing methods are restricted to constrained solutions over limited operator subspaces. Experiments across multiple LLM backbones and pruning strategies demonstrate that our method consistently improves accuracy and perplexity over prior training-free baselines, while preserving the efficiency gains of layer pruning. Official code repository: https://github.com/daniel-eai/ghosted_layers_official_repository/.

cs.LG↗

arXiv2026

Shaping Zero-Shot Coordination via State Blocking

Zero-shot coordination (ZSC) aims to enable agents to cooperate with independently trained partners without prior interaction, a key requirement for real-world multi-agent systems and human-AI collaboration. Existing approaches have largely emphasized increasing partner diversity during training, yet such strategies often fall short of achieving reliable generalization to unseen partners. We introduce State-Blocked Coordination (SBC), a simple yet effective framework that improves ZSC by inducing diverse interaction scenarios without direct environment modification. Specifically, SBC generates a family of virtual environments through state blocking, allowing agents to experience a wide range of suboptimal partner policies. Across multiple benchmarks, SBC demonstrates superior performance in zero-shot coordination, including strong generalization to human partners.

cs.LG↗

arXiv2026

Rethinking Layer Redundancy: Calibration Matters More Than Search in LLM Depth Pruning

Depth pruning improves the inference efficiency of large language models by removing Transformer blocks. Prior work typically treats layer redundancy as an inherent structural property of pretrained networks, emphasizing importance criteria and search algorithms to identify removable layers. In this study, we empirically investigate depth pruning from a functional perspective. Evaluating representative LLM families across diverse calibration configurations and multiple search algorithms, we show that different configurations produce different pruning patterns. Furthermore, under a fixed calibration configuration, complex search algorithms yield marginal performance improvements over simple one-shot methods, converging to similar pruned subsets. Overall, our results suggest that the calibration configuration plays a substantially larger role than the choice of search algorithm in shaping pruning patterns and calibration perplexity, while contributing comparably to variance in downstream reasoning accuracy. This indicates that future pruning efforts may benefit from prioritizing the calibration configuration over search complexity.

cs.LG↗

arXiv2026

Retaining Suboptimal Actions to Follow Shifting Optima in Multi-Agent Reinforcement Learning

Value decomposition is a core approach for cooperative multi-agent reinforcement learning (MARL). However, existing methods still rely on a single optimal action and struggle to adapt when the underlying value function shifts during training, often converging to suboptimal policies. To address this limitation, we propose Successive Sub-value Q-learning (S2Q), which learns multiple sub-value functions to retain alternative high-value actions. Incorporating these sub-value functions into a Softmax-based behavior policy, S2Q encourages persistent exploration and enables $Q^{\text{tot}}$ to adjust quickly to the changing optima. Experiments on challenging MARL benchmarks confirm that S2Q consistently outperforms various MARL algorithms, demonstrating improved adaptability and overall performance. Our code is available at https://github.com/hyeon1996/S2Q.

cs.AI↗

arXiv2026

Clutt3R-Seg: Sparse-view 3D Instance Segmentation for Language-grounded Grasping in Cluttered Scenes

Reliable 3D instance segmentation is fundamental to language-grounded robotic manipulation. Its critical application lies in cluttered environments, where occlusions, limited viewpoints, and noisy masks degrade perception. To address these challenges, we present Clutt3R-Seg, a zero-shot pipeline for robust 3D instance segmentation for language-grounded grasping in cluttered scenes. Our key idea is to introduce a hierarchical instance tree of semantic cues. Unlike prior approaches that attempt to refine noisy masks, our method leverages them as informative cues: through cross-view grouping and conditional substitution, the tree suppresses over- and under-segmentation, yielding view-consistent masks and robust 3D instances. Each instance is enriched with open-vocabulary semantic embeddings, enabling accurate target selection from natural language instructions. To handle scene changes during multi-stage tasks, we further introduce a consistency-aware update that preserves instance correspondences from only a single post-interaction image, allowing efficient adaptation without rescanning. Clutt3R-Seg is evaluated on both synthetic and real-world datasets, and validated on a real robot. Across all settings, it consistently outperforms state-of-the-art baselines in cluttered and sparse-view scenarios. Even on the most challenging heavy-clutter sequences, Clutt3R-Seg achieves an AP@25 of 61.66, over 2.2x higher than baselines, and with only four input views it surpasses MaskClustering with eight views by more than 2x. The code is available at: https://github.com/jeonghonoh/clutt3r-seg.

cs.CV↗

arXiv2026

A.X K1 Technical Report

We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.

cs.CL↗

arXiv2025

Weight Concentration Regularization for Improving Pruning Robustness Under High Sparsity

Deep neural networks achieve outstanding performance across vision and language tasks, yet their large parameter counts limit deployment in resource-constrained settings. One-shot pruning reduces model size without retraining, but models trained with standard objectives often suffer substantial accuracy drops under aggressive sparsity. Prior work mitigates this drop along two directions: regularizers such as $\ell_1$ and DeepHoyer that shape the weight distribution during training, and pruning-robust optimizers such as SAM, CrAM, and S$^2$SAM that flatten the loss landscape. However, existing regularizers either shrink all weights uniformly ($\ell_1$) or induce scale-invariant sparsity (DeepHoyer), without concentrating weight energy onto a small set of informative parameters. We propose a Weight Concentration Regularizer (WCR), a training-time regularizer that amplifies the magnitude of a small subset of parameters while driving the remainder toward zero, so that magnitude pruning predominantly removes parameters with negligible functional contribution. We provide a convergence analysis and evaluate WCR on LLM fine-tuning, image classification, and medical segmentation, demonstrating consistent improvements in pruning robustness across architectures and compatibility with existing pruning-robust optimizers.

cs.LG↗

arXiv2025

Multi-Resolution Model Fusion for Accelerating the Convolutional Neural Network Training

Neural networks are rapidly gaining popularity in scientific research, but training the models is often very time-consuming. Particularly when the training data samples are large high-dimensional arrays, efficient training methodologies that can reduce the computational costs are crucial. To reduce the training cost, we propose a Multi-Resolution Model Fusion (MRMF) method that combines models trained on reduced-resolution data and then refined with data in the original resolution. We demonstrate that these reduced-resolution models and datasets could be generated quickly. More importantly, the proposed approach reduces the training time by speeding up the model convergence in each fusion stage before switching to the final stage of finetuning with data in its original resolution. This strategy ensures the final model retains high-resolution insights while benefiting from the computational efficiency of lower-resolution training. Our experiment results demonstrate that the multi-resolution model fusion method can significantly reduce end-to-end training time while maintaining the same model accuracy. Evaluated using two real-world scientific applications, CosmoFlow and Neuron Inverter, the proposed method improves the training time by up to 47% and 44%, respectively, as compared to the original resolution training, while the model accuracy is not affected.

cs.DC↗

arXiv2025

ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models

Large Vision-Language Models (VLMs) enable strong multimodal reasoning but incur heavy inference costs from redundant visual tokens. Token pruning alleviates this issue, yet existing approaches face limitations. Attention-based methods rely on raw attention scores, which are often unstable across layers and heads and can lead to redundant selections. Diversity-based methods improve robustness by selecting tokens far apart in feature space, but risk dropping regions needed for accurate prediction. We propose ZOO-Prune, a training-free framework built on the intuition that highly sensitive tokens have a stronger influence on the model's output and capture complementary visual cues rather than redundant ones. To achieve this, we estimate token sensitivity using zeroth-order perturbations at the lightweight projection layer. This measures how small random perturbations affect the projected features and enables efficient approximation of each token's influence without backpropagation. Extensive experiments across multiple VLMs and benchmarks show that ZOO-Prune consistently outperforms prior methods while pruning up to 94.4% of tokens without sacrificing accuracy. Our method also improves efficiency, reaching up to 2.30x faster end-to-end inference compared to the baseline.

cs.CV↗

arXiv2025

GEM: A Scale-Aware and Distribution-Sensitive Sparse Fine-Tuning Framework for Effective Downstream Adaptation

Parameter-efficient fine-tuning (PEFT) has become a popular way to adapt large pre-trained models to new tasks. Most PEFT methods update only a small subset of parameters while freezing the rest, avoiding redundant computation. As they maximize the absolute size of the updates without regard to the parameters' original scale, the resulting changes in model behavior can be minimal. In contrast, we maximize updates relative to each parameter's scale, yielding more meaningful downstream adaptation. We propose Gradient-to-Weight Ratio and Entropy-guided Masking (GEM), a parameter scale-aware, distribution-sensitive sparse fine-tuning framework. GEM prioritizes parameters whose updates are significant in proportion to their initial pre-trained values. It also adaptively determines how many parameters to tune at each layer based on the entropy of parameter values, thereby making the most effective use of the computational budget in PEFT. Our empirical study demonstrates the efficacy of GEM on both general-domain tasks (GLUE and SuperGLUE) and domain-specific tasks (GSM8k and MBPP), achieving up to a 1.6% improvement in fine-tuning accuracy over full fine-tuning while updating only 0.1% of model parameters.

cs.LG↗

arXiv2025

Biased Local SGD for Efficient Deep Learning on Heterogeneous Systems

Most parallel neural network training methods assume homogeneous computing resources. For example, synchronous data-parallel SGD suffers from significant synchronization overhead under heterogeneous workloads, often forcing practitioners to rely only on the fastest devices (e.g., GPUs). In this work, we study local SGD for efficient parallel training on heterogeneous systems. We show that intentionally introducing bias in data sampling and model aggregation can effectively harmonize slower CPUs with faster GPUs. Our extensive empirical results demonstrate that a carefully controlled bias significantly accelerates local SGD while achieving comparable or even higher accuracy than synchronous SGD under the same epoch budget. For instance, our method trains ResNet20 on CIFAR-10 with 2 CPUs and 8 GPUs up to 32x faster than synchronous SGD, with nearly identical accuracy. These results provide practical insights into how to flexibly utilize diverse compute resources for deep learning.

cs.LG↗

arXiv2025

Dynamic Rank Adjustment for Accurate and Efficient Neural Network Training

Low-rank training methods reduce the number of trainable parameters by re-parameterizing the weights with matrix decompositions (e.g., singular value decomposition). However, enforcing a fixed low-rank structure caps the rank of the weight matrices and can hinder the model's ability to learn complex patterns. Furthermore, the effective rank of the model's weights tends to decline during training, and this drop is accelerated when the model is reparameterized into a low-rank structure. In this study, we argue that strategically interleaving full-rank training epochs within low-rank training epochs can effectively restore the rank of the model's weights. Based on our findings, we propose a general dynamic-rank training framework that is readily applicable to a wide range of neural-network tasks. We first describe how to adjust the rank of weight matrix to alleviate the inevitable rank collapse that arises during training, and then present extensive empirical results that validate our claims and demonstrate the efficacy of the proposed framework. Our empirical study shows that the proposed method achieves almost the same computational cost as SVD-based low-rank training while achieving a comparable accuracy to full-rank training across various benchmarks.

cs.LG↗

arXiv2025

Focusing Influence Mechanism for Multi-Agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) under sparse rewards remains fundamentally challenging because agents often fail to concentrate their influence, leading to insufficiently coordinated exploration. To address this, we propose the Focusing Influence Mechanism (FIM), a framework that encourages agents to focus their influence on under-explored parts of the state space through an entropy-based criterion, while leveraging eligibility traces to enable multiple agents to consistently align and sustain their influence on the same parts of the state space when beneficial, thereby promoting coordinated and persistent joint behavior. By emphasizing under-explored regions of the state space, FIM facilitates more efficient and structured exploration even under extremely sparse rewards. Across diverse MARL benchmarks, FIM consistently improves cooperative performance over strong baselines.

cs.LG↗

arXiv2025

Layer-wise Adaptive Gradient Norm Penalizing Method for Efficient and Accurate Deep Learning

Sharpness-aware minimization (SAM) is known to improve the generalization performance of neural networks. However, it is not widely used in real-world applications yet due to its expensive model perturbation cost. A few variants of SAM have been proposed to tackle such an issue, but they commonly do not alleviate the cost noticeably. In this paper, we propose a lightweight layer-wise gradient norm penalizing method that tackles the expensive computational cost of SAM while maintaining its superior generalization performance. Our study empirically proves that the gradient norm of the whole model can be effectively suppressed by penalizing the gradient norm of only a few critical layers. We also theoretically show that such a partial model perturbation does not harm the convergence rate of SAM, allowing them to be safely adapted in real-world applications. To demonstrate the efficacy of the proposed method, we perform extensive experiments comparing the proposed method to mini-batch SGD and the conventional SAM using representative computer vision and language modeling benchmarks.

cs.LG↗