arXiv subjects

Xiaoling Zhou

Publications and source records attributed to Xiaoling Zhou.

At least 19 recordsLinked to original sources

arXiv2026

TapeSAC: Automated, Pareto-Optimal Fine-Grained Memory Management for Large Model Training

Selective activation checkpointing (SAC) decides which activations to save for the backward pass and which to recompute: each choice is a point in the plane of peak memory and recomputation time, and the goal is that plane's Pareto frontier. Current practice falls short of this goal. Hand-tuned policies fix a single point that does not adapt when the model, batch or kernels change; PyTorch's SAC reruns each block's forward pass in program order, recomputing values the backward pass never reads; and the memory budget of torch.compile treats recomputation as free or forbidden and selects saved tensors by bytes alone, an assumption that fails for the custom kernels that produce many of a large model's activations. We present TapeSAC, which automates SAC in eager PyTorch without editing the model's forward code. TapeSAC records each block's forward pass at the dispatcher and, during backward, replays only the producers of the values autograd consumes, so saving a tensor eliminates every producer needed only for it. A planner selects the saved tensors from measured recomputation costs under a single limit on step-peak memory. On Qwen3.8-27B and Qwen3.6-35B-A3B, replay alone reduces step time by 3.0% and 1.7% versus full checkpointing at no greater memory cost; memory-aware planning raises these gains to 6.3% and 13.9% under a 90%-of-device-memory limit. The evaluated hand-tuned policies fall short of this memory-recomputation frontier and, on Qwen3.6-35B-A3B, run out of memory at larger batches. Over 500 steps on real data, TapeSAC reduces training time by 8-11%; on Qwen3.5-397B, it is 2.7% faster than a model-specific hand-tuned policy while using 14% less memory.

cs.DC↗

arXiv2026

The Endgame of FSDP Communication: Layout-Aware Collectives on Copy Engines, RDMA, and NVSwitch

Fully Sharded Data Parallelism (FSDP2) maintains parameters and gradients as per-parameter tensors, whereas its collective path operates on flat, rank-major buffers. Converting between these layouts copies all-gather outputs into parameter storage and packs gradients into full-group fp32 buffers. Layout conversion alone accounts for 6.4% of step time in Llama-3.1 70B training at 8K tokens per GPU. Communication kernels impose a second cost: a default NCCL reduce-scatter can slow a concurrent GEMM that would otherwise complete in a single wave by 69%. We present TIDE, a set of zero-copy collectives built on public NCCL 2.31+ APIs. TIDE preserves FSDP2's per-parameter storage and compute kernels while folding layout placement into collective addressing. TIDE-AG delivers shards directly to parameter storage using NVSwitch multicast and paced recursive-doubling RDMA, with no SMs used for transport or placement. TIDE-RS consumes per-parameter bf16 gradients directly and reduces them into fp32 shards (a mixed-precision mode that NCCL does not provide) using copy engines and short-lived thread blocks (CTAs) for node-local accumulation and NVSwitch for cross-node reduction. A two-phase pipeline separates gradient reads from network progress; Slot-Reuse optionally stages gradients in vacated all-gather slots. On 64 GPUs, TIDE reduces step time by up to 27.1% relative to Stock FSDP2 and 6.0% relative to FSDP2+SymmMem across Llama-3.1 70B, Wan2.2, and Qwen3-30B-A3B. It reduces reduce-scatter SM time by factors of 12.5-14.5 and cuts interference with concurrent compute by up to an order of magnitude. Its reduce-scatter buffer remains nearly constant from 16 to 64 GPUs at a fixed per-GPU output size, and its peak memory is up to 4.7 GiB below FSDP2+SymmMem's and, with Slot-Reuse on Wan2.2, below Stock FSDP2's.

cs.DC↗

arXiv2026

Rethinking Positional Encoding and Interaction Granularity for CTR Prediction

Click-through rate (CTR) prediction relies on historical behaviors and contextual features to estimate clicks on a target item. Existing Transformer-based CTR models often adopt index-based or temporal positional encodings and token-level causal interaction. However, we observe that most existing positional encodings do not outperform NoPE, while relaxing causal constraints within target-relative temporal groups also yields gains. These observations motivate a fundamental question: what positional encoding and interaction granularity does CTR prediction actually require? To address this question, we propose \textbf{TRIP}, which combines target-aware rank--time positional encoding with temporal-group interaction modeling. Specifically, Factorized Rank Encoding represents each behavior's ordinal position within the historical sequence through a low-rank factorization, while Target-Relative Time Encoding characterizes its temporal distance to the target on a coarse-to-fine scale. This joint design grounds each behavior's positional representation in its sequence context and temporal relation to the current target. Furthermore, Temporal Group-Causal Interaction enriches intra-group representations through bidirectional interaction and masked reconstruction while preserving inter-group causal order. Finally, behavior representations are condensed into compact group summaries for efficient group-level modeling in subsequent layers. Experiments on four public benchmarks show that TRIP improves AUC by up to 4.45\% relative to the strongest baseline.

cs.IR↗

arXiv2026

Layerwise Repacking: Variable-Length Hybrid-Attention Training without an Extra Redistribution Pass

Variable-length training creates conflicting parallelism requirements across the layers of a hybrid-attention model. The longest samples require wide sequence-parallel (SP) groups of 16 or 32 ranks for full attention, whose cost grows quadratically with sequence length. Sparse-attention layers, dominated by near-linear MoE work, can instead balance most evaluated workloads within node-local SP8 groups. Reusing the wide groups for these layers pushes their all-to-all communication across nodes, where the observed effective rate is 7.0x lower. Yet switching to smaller groups changes which ranks own each sample; a separate redistribution at every full-sparse boundary would introduce another data transfer. Layerwise Repacking keeps each SP16/32 parent for full attention and repacks its samples into node-local sparse children at SP{1, 2, 4, 8}, where hidden states persist between full layers. The layout change rides on the two existing full-layer Ulysses sites: at each, one variable-size all-to-all performs both the head exchange and the sample repacking, so no redistribution pass is added. Because the balance the children can reach depends on which samples share a parent, an auxiliary parent solver rebalances parent membership before the children are formed. On all 22 steps with SP16/32 parents among the first 200 training steps of a 1F3S hybrid-attention model (one full-attention layer per three sparse-attention layers), repacking alone reduces the mean end-to-end time by 5.95% and raises model FLOPs utilization (MFU) by 3.75 percentage points over the baseline with repacking disabled; with the improved solver, end-to-end time falls by 7.61% and MFU rises by 4.83 percentage points.

cs.LG↗

arXiv2026

Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers

Full 3D attention is costly in video Diffusion Transformers. Block-sparse attention reduces this cost by computing only selected attention blocks. However, existing methods either overlook the potential to optimize sparsity allocation or rely on heuristic rules or local metrics to guide it. In this work, we observe that sparsifying attention in different layers and at different denoising steps affects final-video fidelity to substantially different degrees. Yet the identities of the most sensitive layer-step pairs remain nearly unchanged across inputs. Building on these findings, we propose Allocating Layer-step Latency by Output Tests (ALLOT), a training-free sparsity allocation method. Treating each layer-step pair as a cell, ALLOT estimates each cell's contribution to final-video error through (i) cell probing and (ii) trajectory calibration, then jointly optimizes cell densities under a measured-latency budget. It preserves existing selection strategies and kernels, replacing only the allocation component. At equal or shorter wall-clock time, ALLOT outperforms every compared allocation strategy with XAttention, with an 11.8 dB PSNR gain on HunyuanVideo. Extensive experiments across four additional published pipelines on Wan2.1, Wan2.2, and HunyuanVideo show that ALLOT outperforms or ties the default in all nine settings.

cs.CV↗

arXiv2026

KPage: Cross-Block Page Compaction for Hardware-Efficient Sparse Attention

Sparse attention reduces the computational cost of long-sequence video Transformers, but three-dimensional token layouts often leave selected blocks partially occupied, wasting physical tile capacity. We present KPage, a block-sparse attention scheme that packs valid 16-row K/V pages across logical-block boundaries into large tiles directly during loading, without materializing compacted tensors in high-bandwidth memory. KPage also routes Q blocks between 64x256 and 128x128 tiles to reduce query-side padding. Exact masking and K/V-owned page packing preserve the original sparse graph in forward and backward execution. Across five representative workloads, KPage achieves 1.35x-2.06x bare forward-kernel speedup over NVIDIA Block-Sparse Attention (BSA), with a geometric mean of 1.77x. It is near parity for fully occupied blocks (1.03x) but achieves 2.29x geometric-mean speedup on 3D layouts with partial blocks. KPage reduces physical query-key pair capacity from 334.1G to 177.5G while preserving 160.6G useful interactions. Across eight layouts from open-source SSTA, STA, and VSA systems, KPage achieves 1.13x-4.25x forward and 1.02x-2.71x forward-plus-backward speedup over the strongest exact sparse baselines. With independent Q/K partitions, KPage also achieves 1.49x speedup over the strongest exact sparse baseline and 1.57x over BSA on the SVG2 Wan 14B/720p configuration. After compliance review, we will open-source KPage and submit a pull request to NVIDIA cuDNN Frontend.

cs.DC↗

arXiv2026

Generalizable fetal brain biometry via adaptive artificial intelligence

Reliable fetal brain biometry on magnetic resonance imaging (MRI) is essential for diagnosing congenital brain anomalies, but artificial intelligence models often lose accuracy under domain shifts. We developed BioTTA, a backbone-agnostic, source-free unsupervised test-time adaptation framework that adapts pretrained models online to individual unlabeled target volumes. This retrospective multicohort study included 394 reconstructed 3D fetal brain MRI volumes at 20-39 gestational weeks, with 11 biometric measurements annotated by radiologists. BioTTA was developed on single-center turbo spin-echo MRI and evaluated on three datasets representing cross-sequence, cross-center, and cross-nation shifts. Biometry accuracy was assessed using mean absolute error (MAE) and the percentage of correct measurements within 4 mm. A web-based reporting platform mapped measurements onto normative growth trajectories and was evaluated against radiologists in four multicenter disease cohorts covering germinal matrix-intraventricular hemorrhage, ventriculomegaly, and cerebellar abnormality. Across all three target datasets, BioTTA outperformed auto-SVRTK, source-only models, and FetalSynthGen, while matching models finetuned with target labels. MAEs ranged from 1.10 to 1.46 mm with a convolutional backbone and from 1.46 to 2.25 mm with a transformer backbone. Diagnostic accuracy ranged from 82.1% to 95.0%, comparable to a senior radiologist. BioTTA supports accurate, generalizable fetal brain biometry and standardized prenatal brain growth assessment across heterogeneous MRI sources.

cs.CV↗

arXiv2026

DraftAttention2: Fast Video Diffusion with Low-Resolution-Guided Mixed-Precision Attention

Video generation has broad applications in content creation and entertainment. Diffusion transformers have advanced the quality of generated videos, but attention over spatiotemporal tokens becomes increasingly expensive as video resolution and duration increase. We present DraftAttention2, a training-free framework that uses the low-resolution draft attention map to jointly select attention blocks and assign their numerical precision. Specifically, spatial 2D average- and max-pooled queries and keys capture complementary regional statistics to estimate block importance, and a shared ranking assigns higher precision to important blocks, lower precision to less important retained blocks, and skips the rest under configurable budgets. Our analysis separates sparsification error from attention-weighted quantization error, establishing when recovering skipped interactions with low-bit computation tightens the output-error bound. This analysis motivates retaining more interactions at low precision while reserving higher precision for blocks with larger attention mass. To translate these fine-grained assignments into practical speedups, we further develop fused operand preparation and a single attention kernel with precision-specific phases, sharing data movement, softmax statistics, and output accumulation across precisions. Experiments demonstrate that our method achieves a superior quality-efficiency trade-off over existing efficient video generation methods. Notably, its advantage is particularly pronounced for few-step video diffusion, where jointly combining sparsity with 4- and 8-bit mixed-precision computation substantially improves generation quality while retaining significant acceleration. Code is available at https://github.com/anemoi-project/anemoi

cs.CV↗

arXiv2026

Annotation-free deep learning for detection and segmentation of fetal germinal matrix-intraventricular hemorrhage in brain MRI

Prenatal germinal matrix-intraventricular hemorrhage (GMH-IVH) is a leading cause of infant mortality and neurodevelopmental impairment, yet its manual diagnosis and lesion segmentation on fetal brain MRI are labor-intensive and error-prone. Although supervised deep learning offers potential for automation, it typically requires large amounts of annotated GMH-IVH data, which are challenging to obtain for such a rare condition (0.5-0.9 per 1000 pregnancies). To address these problems, an annotation-free deep learning framework, FreeHemoSeg, was developed for automated detection and segmentation of GMH-IVH without any real patient annotations. Instead of learning from expert labels, FreeHemoSeg was trained on pseudo GMH-IVH images synthesized from normal fetal data guided by medical priors. The framework was evaluated in a retrospective multicentre study of 1,674 stacks of 2D T2-weighted MRI from 558 pregnant women, using data from one hospital for internal training and validation and two hospitals for external validation. FreeHemoSeg achieved the highest diagnostic and segmentation performance in both internal validation (AUROC: 0.959; AUPR: 0.928; sensitivity: 0.914; specificity: 0.966; DSC: 0.559) and external validation (AUROC: 0.930; AUPR: 0.884; sensitivity: 0.824; specificity: 0.943; DSC: 0.512), outperforming a supervised model trained on limited empirical data and unsupervised anomaly detection methods. Moreover, FreeHemoSeg assistance improved radiologists' sensitivity (from 0.882 to 0.941-1.000) and diagnostic confidence, while reducing interpretation time by 16.0-52.7%. We anticipate its immediate utility in supporting earlier diagnosis, prognostic counselling, and perinatal planning for fetal GMH-IVH. Code: https://github.com/Arktis2022/FreeHemoSeg.

eess.IV↗

arXiv2026

Mind Dreamer: Untethering Imagination via Active Causal Intervention on Latent Manifolds

Model-Based Reinforcement Learning yields sample efficiency via latent imagination, yet remains constrained by Historical Tethering: imagination is typically initialized from observed states. This creates a learning asymmetry, where the world model's manifold discovery outpaces the policy's sparse-reward optimization. We propose Mind Dreamer (MD), a framework that instantiates Active Causal Intervention to transcend Markovian continuity. MD reformulates discovery as the minimization of a global Relay Expected Free Energy. Instead of initializing from historical data, it draws initial states from an adversarial generator $s_0 \sim p_{gen}(\cdot)$, creating non-continuous latent jumps to epistemic blind spots that are physically plausible yet cognitively challenging. We derive Relay Value Function and Relay Uncertainty Function to resolve the credit assignment paradox across these spatial ruptures. Treating synthesized anchors as interventional intermediary states, these potentials propagate pragmatic and epistemic value through Bellman-style backups. Notably, we prove that uncertainty propagation across discontinuities necessitates a quadratic discount $γ^2$, establishing a formal epistemic horizon. Theoretically, MD approximates a variance-minimizing importance sampler that expands the manifold's spectral gap, reducing the hitting time to critical bottleneck states. Empirically, MD achieves a 1.67$\times$ average speedup over DreamerV3 on DeepMind Control Suite, reaching 8.8$\times$ in sparse-reward tasks.

cs.LG↗

arXiv2025

Data Valuation by Fusing Global and Local Statistical Information

Data valuation has garnered increasing attention in recent years, given the critical role of high-quality data in various applications. Among diverse data valuation approaches, Shapley value-based methods are predominant due to their strong theoretical grounding. However, the exact computation of Shapley values is often computationally prohibitive, prompting the development of numerous approximation techniques. Despite notable advancements, existing methods generally neglect the incorporation of value distribution information and fail to account for dynamic data conditions, thereby compromising their performance and application potential. In this paper, we highlight the crucial role of both global and local statistical properties of value distributions in the context of data valuation for machine learning. First, we conduct a comprehensive analysis of these distributions across various simulated and real-world datasets, uncovering valuable insights and key patterns. Second, we propose an enhanced data valuation method that fuses the explored distribution characteristics into two regularization terms to refine Shapley value estimation. The proposed regularizers can be seamlessly incorporated into various existing data valuation methods. Third, we introduce a novel approach for dynamic data valuation that infers updated data values without recomputing Shapley values, thereby significantly improving computational efficiency. Extensive experiments have been conducted across a range of tasks, including Shapley value estimation, value-based data addition and removal, mislabeled data detection, and dynamic data valuation. The results showcase the consistent effectiveness and efficiency of our proposed methodologies, affirming the significant potential of global and local value distributions in data valuation.

cs.LG↗

arXiv2025

Decoupling Safety into Orthogonal Subspace: Cost-Efficient and Performance-Preserving Alignment for Large Language Models

Safety alignment is essential for building trustworthy artificial intelligence, yet it remains challenging to enhance model safety without degrading general performance. Current approaches require computationally expensive searches for the optimal proportion of safety-critical and general-purpose data to balance safety and general performance, incurring high costs with limited gains. In this work, we show that LoRA-based Refusal-training enables performance-preserving safety alignment even when trained solely on safety data, demonstrating that LoRA serves as cost-efficient, performance-preserving, and plug-and-play safety patches. Beyond empirical findings, we provide both theoretical and experimental evidence that LoRA effectively decouples safety into a low-rank subspace largely orthogonal to the model's intrinsic transformation space, ensuring that safety enhancements do not interfere with inherent capabilities.

cs.CL↗

arXiv2025

Seed&Steer: Guiding Large Language Models with Compilable Prefix and Branch Signals for Unit Test Generation

Unit tests play a vital role in the software development lifecycle. Recent advances in Large Language Model (LLM)-based approaches have significantly improved automated test generation, garnering attention from both academia and industry. We revisit LLM-based unit test generation from a novel perspective by decoupling prefix generation and assertion generation. To characterize their respective challenges, we define Initialization Complexity and adopt Cyclomatic Complexity to measure the difficulty of prefix and assertion generation, revealing that the former primarily affects compilation success, while the latter influences test coverage. To address these challenges, we propose Seed&Steer, a two-step approach that combines traditional unit testing techniques with the capabilities of large language models. Seed&Steer leverages conventional unit testing tools (e.g., EvoSuite) to generate method invocations with high compilation success rates, which serve as seeds to guide LLMs in constructing effective test contexts. It then introduces branching cues to help LLMs explore diverse execution paths (e.g., normal, boundary, and exception cases) and generate assertions with high coverage. We evaluate Seed&Steer on five real-world Java projects against state-of-the-art baselines. Results show that Seed&Steer improves the compilation pass rate by approximately 7%, successfully compiling 792 and 887 previously failing cases on two LLMs. It also achieves up to ~73% branch and line coverage across focal methods of varying complexity, with coverage improvements ranging from 1.09* to 1.26*. Our code, dataset, and experimental scripts will be publicly released to support future research and reproducibility.

cs.SE↗

arXiv2025

Implicit Counterfactual Data Augmentation for Robust Learning

Machine learning models are prone to capturing the spurious correlations between non-causal attributes and classes, with counterfactual data augmentation being a promising direction for breaking these spurious associations. However, generating counterfactual data explicitly poses a challenge, and incorporating augmented data into the training process decreases training efficiency. This study proposes an Implicit Counterfactual Data Augmentation (ICDA) method to remove spurious correlations and make stable predictions. Specifically, first, a novel sample-wise augmentation strategy is developed that generates semantically and counterfactually meaningful deep features with distinct augmentation strength for each sample. Second, we derive an easy-to-compute surrogate loss on the augmented feature set when the number of augmented samples becomes infinite. Third, two concrete schemes are proposed, including direct quantification and meta-learning, to derive the key parameters for the robust loss. In addition, ICDA is explained from a regularization perspective, revealing its capacity to improve intra-class compactness and augment margins at both class and sample levels. Extensive experiments have been conducted across various biased learning scenarios covering both image and text datasets, demonstrating that ICDA consistently enhances the generalization and robustness performance of popular networks.

cs.LG↗

arXiv2025

Mitigating Spurious Correlations with Causal Logit Perturbation

Deep learning has seen widespread success in various domains such as science, industry, and society. However, it is acknowledged that certain approaches suffer from non-robustness, relying on spurious correlations for predictions. Addressing these limitations is of paramount importance, necessitating the development of methods that can disentangle spurious correlations. {This study attempts to implement causal models via logit perturbations and introduces a novel Causal Logit Perturbation (CLP) framework to train classifiers with generated causal logit perturbations for individual samples, thereby mitigating the spurious associations between non-causal attributes (i.e., image backgrounds) and classes.} {Our framework employs a} perturbation network to generate sample-wise logit perturbations using a series of training characteristics of samples as inputs. The whole framework is optimized by an online meta-learning-based learning algorithm and leverages human causal knowledge by augmenting metadata in both counterfactual and factual manners. Empirical evaluations on four typical biased learning scenarios, including long-tail learning, noisy label learning, generalized long-tail learning, and subpopulation shift learning, demonstrate that CLP consistently achieves state-of-the-art performance. Moreover, visualization results support the effectiveness of the generated causal perturbations in redirecting model attention towards causal image attributes and dismantling spurious associations.

cs.LG↗

arXiv2024

Enhancing In-Context Learning via Implicit Demonstration Augmentation

The emergence of in-context learning (ICL) enables large pre-trained language models (PLMs) to make predictions for unseen inputs without updating parameters. Despite its potential, ICL's effectiveness heavily relies on the quality, quantity, and permutation of demonstrations, commonly leading to suboptimal and unstable performance. In this paper, we tackle this challenge for the first time from the perspective of demonstration augmentation. Specifically, we start with enriching representations of demonstrations by leveraging their deep feature distribution. We then theoretically reveal that when the number of augmented copies approaches infinity, the augmentation is approximately equal to a novel logit calibration mechanism integrated with specific statistical properties. This insight results in a simple yet highly efficient method that significantly improves the average and worst-case accuracy across diverse PLMs and tasks. Moreover, our method effectively reduces performance variance among varying demonstrations, permutations, and templates, and displays the capability to address imbalanced class distributions.

cs.LG↗

arXiv2024

Boosting Model Resilience via Implicit Adversarial Data Augmentation

Data augmentation plays a pivotal role in enhancing and diversifying training data. Nonetheless, consistently improving model performance in varied learning scenarios, especially those with inherent data biases, remains challenging. To address this, we propose to augment the deep features of samples by incorporating their adversarial and anti-adversarial perturbation distributions, enabling adaptive adjustment in the learning difficulty tailored to each sample's specific characteristics. We then theoretically reveal that our augmentation process approximates the optimization of a surrogate loss function as the number of augmented copies increases indefinitely. This insight leads us to develop a meta-learning-based framework for optimizing classifiers with this novel loss, introducing the effects of augmentation while bypassing the explicit augmentation process. We conduct extensive experiments across four common biased learning scenarios: long-tail learning, generalized long-tail learning, noisy label learning, and subpopulation shift learning. The empirical results demonstrate that our method consistently achieves state-of-the-art performance, highlighting its broad adaptability.

cs.LG↗

arXiv2023

Combining Adversaries with Anti-adversaries in Training

Adversarial training is an effective learning technique to improve the robustness of deep neural networks. In this study, the influence of adversarial training on deep learning models in terms of fairness, robustness, and generalization is theoretically investigated under more general perturbation scope that different samples can have different perturbation directions (the adversarial and anti-adversarial directions) and varied perturbation bounds. Our theoretical explorations suggest that the combination of adversaries and anti-adversaries (samples with anti-adversarial perturbations) in training can be more effective in achieving better fairness between classes and a better tradeoff between robustness and generalization in some typical learning scenarios (e.g., noisy label learning and imbalance learning) compared with standard adversarial training. On the basis of our theoretical findings, a more general learning objective that combines adversaries and anti-adversaries with varied bounds on each training sample is presented. Meta learning is utilized to optimize the combination weights. Experiments on benchmark datasets under different learning scenarios verify our theoretical findings and the effectiveness of the proposed methodology.

cs.LG↗