arXiv subjects

Vincent-Daniel Yun

Publications and source records attributed to Vincent-Daniel Yun.

12 recordsLinked to original sources

arXiv2026

Forward-Free LLM Depth Pruning via Weight Redundancy

Depth pruning reduces large language model (LLM) inference cost by removing complete Transformer blocks. Activation-based methods collect hidden states through forward passes on calibration data, while existing forward-free methods score each Transformer block separately without measuring similarity between blocks. We propose Weight-Redundancy Pruning (WRP), a forward-free depth-pruning method that estimates inter-layer redundancy from checkpoint weights to select blocks without calibration data or model forward passes. WRP compares attention output and MLP down-projection weights across layers and combines their pairwise similarities with relative projection-scale information. The resulting all-pairs similarity matrix guides layer grouping and block selection. Across multiple pruning settings, model families, and downstream tasks, WRP consistently outperforms existing forward-free magnitude pruning and approaches the performance of activation-based methods.

cs.LG↗

arXiv2026

KV-Rescue: Recovering Reasoning Language Model KV Eviction Loss via Stepwise Interleaving

KV-cache eviction caps the memory cost of long reasoning traces but is inherently lossy because the model decodes from a partial view of its history. Under aggressive budgets, this not only lowers accuracy but can also cause runaway degeneration, where the model produces incoherent or repetitive tokens until reaching the length limit. We characterize much of this loss as an information gapf caused by missing context, rather than a capability gap caused by limited model capacity. An evicted 7B model and a full-context 1.5B model make complementary errors, and an oracle choice between their answers recovers 79% of the accuracy gap to the full-KV 7B model. Based on this observation, we propose KV-Rescue, a training-free inference framework that bridges the information gap introduced by KV eviction using a lightweight full-context helper. KV-Rescue interleaves reasoning steps from the two models into a shared trajectory. An online detector uses entropy and compressibility to terminate the generation of incoherent or repetitive base-model candidates early. Across five math benchmarks with Qwen2.5-Math 7B and 72B, KV-Rescue recovers an average of 87% of the accuracy lost to eviction at eviction budget B=64. A decode-cost analysis further shows that preventing runaway degeneration cuts base-model token generation by 43% on average.

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

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

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

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

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↗

arXiv2026

Robust Multi-Agent LLMs under Byzantine Faults

Large language model (LLM) agents increasingly collaborate over peer-to-peer networks to improve their reliability. However, these same interactions can also introduce vulnerability to unreliable or Byzantine agents that can propagate incorrect information and degrade overall system performance. To address this, we propose Self-Anchored Consensus (SAC), a fully decentralized filter-and-refine protocol in which agents iteratively exchange responses, locally evaluate and filter unreliable messages, and refine their own outputs. We present $(F{+}1)$-robustness conditions on the communication graph that ensure honest agents preserve and propagate reliable information despite Byzantine influence. Experiments across diverse open- and closed-weight LLMs on mathematical and commonsense reasoning benchmarks show that SAC effectively suppresses Byzantine influence and consistently improves performance across diverse communication topologies, whereas prior methods degrade significantly under Byzantine attacks.

cs.MA↗

arXiv2026

Sharpness-Aware Minimization with Z-Score Gradient Filtering

Deep neural networks achieve high performance across many domains but can still face challenges in generalization when optimization is influenced by small or noisy gradient components. Sharpness-Aware Minimization improves generalization by perturbing parameters toward directions of high curvature, but it uses the entire gradient vector, which means that small or noisy components may affect the ascent step and cause the optimizer to miss optimal solutions. We propose Z-Score Filtered Sharpness-Aware Minimization, which applies Z-score based filtering to gradients in each layer. Instead of using all gradient components, a mask is constructed to retain only the top percentile with the largest absolute Z-scores. The percentile threshold $Q_p$ determines how many components are kept, so that the ascent step focuses on directions that stand out most compared to the average of the layer. This selective perturbation refines the search toward flatter minima while reducing the influence of less significant gradients. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with architectures including ResNet, VGG, and Vision Transformers show that the proposed method consistently improves test accuracy compared to Sharpness-Aware Minimization and its variants. The code repository is available at: https://github.com/YUNBLAK/Sharpness-Aware-Minimization-with-Z-Score-Gradient-Filtering

cs.LG↗

arXiv2026

Why Does Stochastic Gradient Descent Slow Down in Low-Precision Training?

Low-precision training has become crucial for reducing the computational and memory costs of large-scale deep learning. However, quantizing gradients introduces magnitude shrinkage, which can change how stochastic gradient descent (SGD) converges. In this study, we explore SGD convergence under a gradient shrinkage model, where each stochastic gradient is scaled by a factor $ q_k \in (0,1] $. We show that this shrinkage affect the usual stepsize $ μ_k $ with an effective stepsize $ μ_k q_k $, slowing convergence when $ q_{\min} < 1 $. With typical smoothness and bounded-variance assumptions, we prove that low-precision SGD still converges, but at a slower pace set by $ q_{\min} $, and with a higher steady error level due to quantization effects. We analyze theoretically how lower numerical precision slows training by treating it as gradient shrinkage within the standard SGD convergence setup.

cs.LG↗

arXiv2025

MedCLM: Learning to Localize and Reason via a CoT-Curriculum in Medical Vision-Language Models

Bridging clinical diagnostic reasoning with AI remains a central challenge in medical imaging. We introduce MedCLM, an automated pipeline that converts detection datasets into large-scale medical visual question answering (VQA) data with Chain-of-Thought (CoT) reasoning by linking lesion boxes to organ segmentation and structured rationales. These contextual signals enable medical vision-language models to generate question-answer pairs with step-by-step reasoning. To utilize this data effectively, we propose an Integrated CoT-Curriculum Strategy composed of an Easy stage with explicit lesion boxes for visual grounding, a Medium stage that encourages implicit localization, and a Hard stage for weakly supervised reasoning. Experimental results demonstrate that MedCLM attains state-of-the-art performance on several medical VQA benchmarks, providing a scalable framework for developing clinically aligned medical vision-language models.

cs.CV↗

arXiv2025

Insights from Gradient Dynamics: Gradient Autoscaled Normalization

Gradient dynamics play a central role in determining the stability and generalization of deep neural networks. In this work, we provide an empirical analysis of how variance and standard deviation of gradients evolve during training, showing consistent changes across layers and at the global scale in convolutional networks. Motivated by these observations, we propose a hyperparameter-free gradient normalization method that aligns gradient scaling with their natural evolution. This approach prevents unintended amplification, stabilizes optimization, and preserves convergence guarantees. Experiments on the challenging CIFAR-100 benchmark with ResNet-20, ResNet-56, and VGG-16-BN demonstrate that our method maintains or improves test accuracy even under strong generalization. Beyond practical performance, our study highlights the importance of directly tracking gradient dynamics, aiming to bridge the gap between theoretical expectations and empirical behaviors, and to provide insights for future optimization research.

cs.LG↗