arXiv subjects

Yang Hu

Publications and source records attributed to Yang Hu.

At least 73 records · Page 4Linked to original sources

arXiv2025

Optimal Model Selection for Conformalized Robust Optimization

In decision-making under uncertainty, Contextual Robust Optimization (CRO) provides reliability by minimizing the worst-case decision loss over a prediction set. While recent advances use conformal prediction to construct prediction sets for machine learning models, the downstream decisions critically depend on model selection. This paper introduces novel model selection frameworks for CRO that unify robustness control with decision risk minimization. We first propose Conformalized Robust Optimization with Model Selection (CROMS), a framework that selects the model to approximately minimize the averaged decision risk in CRO solutions. Given the target robustness level 1-α, we present a computationally efficient algorithm called E-CROMS, which achieves asymptotic robustness control and decision optimality. To correct the control bias in finite samples, we further develop two algorithms: F-CROMS, which ensures a 1-αrobustness but requires searching the label space; and J-CROMS, which offers lower computational cost while achieving a 1-2αrobustness. Furthermore, we extend the CROMS framework to the individualized setting, where model selection is performed by minimizing the conditional decision risk given the covariates of the test data. This framework advances conformal prediction methodology by enabling covariate-aware model selection. Numerical results demonstrate significant improvements in decision efficiency across diverse synthetic and real-world applications, outperforming baseline approaches.

stat.ML↗

arXiv2025

Designing Spatial Architectures for Sparse Attention: STAR Accelerator via Cross-Stage Tiling

Large language models (LLMs) rely on self-attention for contextual understanding, demanding high-throughput inference and large-scale token parallelism (LTPP). Existing dynamic sparsity accelerators falter under LTPP scenarios due to stage-isolated optimizations. Revisiting the end-to-end sparsity acceleration flow, we identify an overlooked opportunity: cross-stage coordination can substantially reduce redundant computation and memory access. We propose STAR, a cross-stage compute- and memory-efficient algorithm-hardware co-design tailored for Transformer inference under LTPP. STAR introduces a leading-zero-based sparsity prediction using log-domain add-only operations to minimize prediction overhead. It further employs distributed sorting and a sorted updating FlashAttention mechanism, guided by a coordinated tiling strategy that enables fine-grained stage interaction for improved memory efficiency and latency. These optimizations are supported by a dedicated STAR accelerator architecture, achieving up to 9.2$\times$ speedup and 71.2$\times$ energy efficiency over A100, and surpassing SOTA accelerators by up to 16.1$\times$ energy and 27.1$\times$ area efficiency gains. Further, we deploy STAR onto a multi-core spatial architecture, optimizing dataflow and execution orchestration for ultra-long sequence processing. Architectural evaluation shows that, compared to the baseline design, Spatial-STAR achieves a 20.1$\times$ throughput improvement.

cs.AR↗

arXiv2025

cuPilot: A Strategy-Coordinated Multi-agent Framework for CUDA Kernel Evolution

Optimizing CUDA kernels is a challenging and labor-intensive task, given the need for hardware-software co-design expertise and the proprietary nature of high-performance kernel libraries. While recent large language models (LLMs) combined with evolutionary algorithms show promise in automatic kernel optimization, existing approaches often fall short in performance due to their suboptimal agent designs and mismatched evolution representations. This work identifies these mismatches and proposes cuPilot, a strategy-coordinated multi-agent framework that introduces strategy as an intermediate semantic representation for kernel evolution. Key contributions include a strategy-coordinated evolution algorithm, roofline-guided prompting, and strategy-level population initialization. Experimental results show that the generated kernels by cuPilot achieve an average speed up of 3.09$\times$ over PyTorch on a benchmark of 100 kernels. On the GEMM tasks, cuPilot showcases sophisticated optimizations and achieves high utilization of critical hardware units. The generated kernels are open-sourced at https://github.com/champloo2878/cuPilot-Kernels.git.

cs.AI↗

arXiv2025

TEMP: A Memory Efficient Physical-aware Tensor Partition-Mapping Framework on Wafer-scale Chips

Large language models (LLMs) demand significant memory and computation resources. Wafer-scale chips (WSCs) provide high computation power and die-to-die (D2D) bandwidth but face a unique trade-off between on-chip memory and compute resources due to limited wafer area. Therefore, tensor parallelism strategies for wafer should leverage communication advantages while maintaining memory efficiency to maximize WSC performance. However, existing approaches fail to address these challenges. To address these challenges, we propose the tensor stream partition paradigm (TSPP), which reveals an opportunity to leverage WSCs' abundant communication bandwidth to alleviate stringent on-chip memory constraints. However, the 2D mesh topology of WSCs lacks long-distance and flexible interconnects, leading to three challenges: 1) severe tail latency, 2) prohibitive D2D traffic contention, and 3) intractable search time for optimal design. We present TEMP, a framework for LLM training on WSCs that leverages topology-aware tensor-stream partition, traffic-conscious mapping, and dual-level wafer solving to overcome hardware constraints and parallelism challenges. These integrated approaches optimize memory efficiency and throughput, unlocking TSPP's full potential on WSCs. Evaluations show TEMP achieves 1.7x average throughput improvement over state-of-the-art LLM training systems across various models.

cs.AR↗

arXiv2025

WATOS: Efficient LLM Training Strategies and Architecture Co-exploration for Wafer-scale Chip

Training large language models (LLMs) imposes extreme demands on computation, memory capacity, and interconnect bandwidth, driven by their ever-increasing parameter scales and intensive data movement. Wafer-scale integration offers a promising solution by densely integrating multiple single-die chips with high-speed die-to-die (D2D) interconnects. However, the limited wafer area necessitates trade-offs among compute, memory, and communication resources. Fully harnessing the potential of wafer-scale integration while mitigating its architectural constraints is essential for maximizing LLM training performance. This imposes significant challenges for the co-optimization of architecture and training strategies. Unfortunately, existing approaches all fall short in addressing these challenges. To bridge the gap, we propose WATOS, a co-exploration framework for LLM training strategy and wafer-scale architecture. We first define a highly configurable hardware template designed to explore optimal architectural parameters for wafer-scale chips. Based on it, we capitalize on the high D2D bandwidth and fine-grained operation advantages inherent to wafer-scale chips to explore optimal parallelism and resource allocation strategies, effectively addressing the memory underutilization issues during LLM training. Compared to the state-of-the-art (SOTA) LLM training framework Megatron and Cerebras' weight streaming wafer training strategy, WATOS can achieve an average overall throughput improvement of 2.74x and 1.53x across various LLM models, respectively. In addition, we leverage WATOS to reveal intriguing insights about wafer-scale architecture design with the training of LLM workloads.

eess.SP↗

arXiv2025

Casimir radiation with Weyl semimetals

When Casimir friction torque acts upon a rotated nanoparticle (NP), mechanical energy can be transformed into thermal energy, known as Casimir radiation, which significantly affects the thermal performance of nanoelectromechanical systems. In this work, we investigate Casimir radiation with nonreciprocal Weyl semimetals (WSM) NP levitated on a plate. WSM NP with inherent nonreciprocity has a radiative heat flux 27 times higher than NP with degenerate modes. The underlying physics is elucidated by the coupling and decoupling of the electromagnetic local density of states between nonreciprocal WSN NP and the plate in the near-field. The three-fold localized plasmon modes of WSM NP split into localized circular modes with strong gyrotropic response, which opens up new channels for Casimir radiation. This work provides a new method for nanoscale energy conversion in NP systems.

physics.optics↗

arXiv2025

BitStopper: An Efficient Transformer Attention Accelerator via Stage-fusion and Early Termination

Attention-based large language models (LLMs) have transformed modern AI applications, but the quadratic cost of self-attention imposes significant compute and memory overhead. Dynamic sparsity (DS) attention mitigates this, yet its hardware efficiency is limited by the added prediction stage and the heavy memory traffic it entails. To address these limitations, this paper proposes BitStopper, a fine-grained algorithm-architecture co-design that operates without a sparsity predictor. First, a bit-serial enable stage fusion (BESF) mechanism is proposed to reuse and minimize the memory access by progressively terminating trivial tokens and merging the prediction stage into the execution stage. Second, a lightweight and adaptive token selection (LATS) strategy is developed to work in concert with the bit-level sparsity speculation. Third, a bit-level asynchronous processing (BAP) strategy is employed to improve compute utilization during the on-demand bit-grained memory fetching. Finally, an elaborate architecture is designed to translate the theoretical complexity reduction into practical performance improvement. Extensive evaluations demonstrate that, compared to state-of-the-art (SOTA) Transformer accelerators, BitStopper achieves 2.03x and 1.89x speedups over Sanger and SOFA, respectively, while delivering 2.4x and 2.1x improvements in energy efficiency.

cs.LG↗

arXiv2025

TraceTrans: Translation and Spatial Tracing for Surgical Prediction

Image-to-image translation models have achieved notable success in converting images across visual domains and are increasingly used for medical tasks such as predicting post-operative outcomes and modeling disease progression. However, most existing methods primarily aim to match the target distribution and often neglect spatial correspondences between the source and translated images. This limitation can lead to structural inconsistencies and hallucinations, undermining the reliability and interpretability of the predictions. These challenges are accentuated in clinical applications by the stringent requirement for anatomical accuracy. In this work, we present TraceTrans, a novel deformable image translation model designed for post-operative prediction that generates images aligned with the target distribution while explicitly revealing spatial correspondences with the pre-operative input. The framework employs an encoder for feature extraction and dual decoders for predicting spatial deformations and synthesizing the translated image. The predicted deformation field imposes spatial constraints on the generated output, ensuring anatomical consistency with the source. Extensive experiments on medical cosmetology and brain MRI datasets demonstrate that TraceTrans delivers accurate and interpretable post-operative predictions, highlighting its potential for reliable clinical deployment.

eess.IV↗

arXiv2025

LAPA: Log-Domain Prediction-Driven Dynamic Sparsity Accelerator for Transformer Model

Attention-based Transformers have revolutionized natural language processing (NLP) and shown strong performance in computer vision (CV) tasks. However, as the input sequence varies, the computational bottlenecks in Transformer models exhibit dynamic behavior across stages, which calls for a cross-stage sparse acceleration strategy. Unfortunately, most existing sparse Transformer approaches are single-stage based, and their sparsity prediction mechanisms lead to significant power overhead when applied across multiple stages. To this end, this paper proposes a log-domain attention prediction algorithm-architecture co-design, named LAPA. First, an asymmetric leading one computing (ALOC) scheme is designed to eliminate expensive multiplications. Next, a mixed-precision multi-round shifting accumulation (MRSA) mechanism is further proposed to mitigate the accumulation overhead. A data-feature dependent filter (DDF) strategy is designed to work in concert with the MRSA process. Finally, an elaborate accelerator is designed to translate the theoretical enhancement into practical hardware improvement. Experimental results show that LAPA achieves 3.52x, 3.24x and 2.79x higher energy efficiency than the state-of-the-art (SOTA) works Spatten, Sanger and FACT, respectively.

cs.LG↗

arXiv2025

New Algebrization Barriers to Circuit Lower Bounds via Communication Complexity of Missing-String

The *algebrization barrier*, proposed by Aaronson and Wigderson (STOC '08, ToCT '09), captures the limitations of many complexity-theoretic techniques based on arithmetization. Notably, several circuit lower bounds that overcome the relativization barrier (Buhrman--Fortnow--Thierauf, CCC '98; Vinodchandran, TCS '05; Santhanam, STOC '07, SICOMP '09) remain subject to the algebrization barrier. In this work, we establish several new algebrization barriers to circuit lower bounds by studying the communication complexity of the following problem, called XOR-Missing-String: For $m < 2^{n/2}$, Alice gets a list of $m$ strings $x_1, \dots, x_m\in\{0, 1\}^n$, Bob gets a list of $m$ strings $y_1, \dots, y_m\in\{0, 1\}^n$, and the goal is to output a string $s\in\{0, 1\}^n$ that is not equal to $x_i\oplus y_j$ for any $i, j\in [m]$. 1. We construct an oracle $A_1$ and its multilinear extension $\widetilde{A_1}$ such that ${\sf PostBPE}^{\widetilde{A_1}}$ has linear-size $A_1$-oracle circuits on infinitely many input lengths. 2. We construct an oracle $A_2$ and its multilinear extension $\widetilde{A_2}$ such that ${\sf BPE}^{\widetilde{A_2}}$ has linear-size $A_2$-oracle circuits on all input lengths. 3. Finally, we study algebrization barriers to circuit lower bounds for $\sf MA_E$. Buhrman, Fortnow, and Thierauf proved a *sub-half-exponential* circuit lower bound for $\sf MA_E$ via algebrizing techniques. Toward understanding whether the half-exponential bound can be improved, we define a natural subclass of $\sf MA_E$ that includes their hard $\sf MA_E$ language, and prove the following result: For every *super-half-exponential* function $h(n)$, we construct an oracle $A_3$ and its multilinear extension $\widetilde{A_3}$ such that this natural subclass of ${\sf MA}_{\sf E}^{\widetilde{A_3}}$ has $h(n)$-size $A_3$-oracle circuits on all input lengths.

cs.CC↗

arXiv2025

Histology-informed tiling of whole tissue sections improves the interpretability and predictability of cancer relapse and genetic alterations

Histopathologists establish cancer grade by assessing histological structures, such as glands in prostate cancer. Yet, digital pathology pipelines often rely on grid-based tiling that ignores tissue architecture. This introduces irrelevant information and limits interpretability. We introduce histology-informed tiling (HIT), which uses semantic segmentation to extract glands from whole slide images (WSIs) as biologically meaningful input patches for multiple-instance learning (MIL) and phenotyping. Trained on 137 samples from the ProMPT cohort, HIT achieved a gland-level Dice score of 0.83 +/- 0.17. By extracting 380,000 glands from 760 WSIs across ICGC-C and TCGA-PRAD cohorts, HIT improved MIL models AUCs by 10% for detecting copy number variation (CNVs) in genes related to epithelial-mesenchymal transitions (EMT) and MYC, and revealed 15 gland clusters, several of which were associated with cancer relapse, oncogenic mutations, and high Gleason. Therefore, HIT improved the accuracy and interpretability of MIL predictions, while streamlining computations by focussing on biologically meaningful structures during feature extraction.

cs.CV↗

arXiv2025

Nash Social Welfare with Submodular Valuations: Approximation Algorithms and Integrality Gaps

We study the problem of allocating items to agents with submodular valuations with the goal of maximizing the weighted Nash social welfare (NSW). The best-known results for unweighted and weighted objectives are the $(4+ε)$ approximation given by Garg, Husic, Li, Végh, and Vondrák~[STOC 2023] and the $(233+ε)$ approximation given by Feng, Hu, Li, and Zhang~[STOC 2025], respectively. In this work, we present a $(3.56+ε)$-approximation algorithm for weighted NSW maximization with submodular valuations, simultaneously improving the previous approximation ratios of both the weighted and unweighted NSW problems. Our algorithm solves the configuration LP of Feng, Hu, Li, and Zhang~[STOC 2025] via a stronger separation oracle that loses an $e/(e-1)$ factor only on small items, and then rounds the solution via a new bipartite multigraph construction. Some key technical ingredients of our analysis include a greedy proxy function, additive within each configuration, that preserves the LP value while lower-bounding the rounded solution, together with refined concentration bounds and a series of mathematical programs analyzed partly by computer assistance. On the hardness side, we prove that the configuration LP for weighted NSW with submodular valuations has an integrality gap of at least $(2^{\ln 2}-ε) \approx 1.617 - ε$, which is larger than the current best-known $e/(e-1)-ε\approx 1.582-ε$ hardness~[SODA 2020]. For additive valuations, we show an integrality gap of $(e^{1/e}-ε)$, which proves the tightness of the approximation ratio in~[ICALP 2024] for algorithms based on the configuration LP. For unweighted NSW with additive valuations, we show an integrality gap of $(2^{1/4}-ε) \approx 1.189-ε$, again larger than the current best-known $\sqrt{8/7} \approx 1.069$-hardness~[Math. Oper. Res. 2024].

cs.GT↗

arXiv2025

Constant Approximation for Weighted Nash Social Welfare with Submodular Valuations

We study the problem of assigning items to agents so as to maximize the \emph{weighted} Nash Social Welfare (NSW) under submodular valuations. The best-known result for the problem is an $O(nw_{\max})$-approximation due to Garg, Husic, Li, Végh, and Vondrák~[STOC 2023], where $w_{\max}$ is the maximum weight over all agents. Obtaining a constant approximation algorithm is an open problem in the field that has recently attracted considerable attention. We give the first such algorithm for the problem, thus solving the open problem in the affirmative. Our algorithm is based on the natural Configuration LP for the problem, which was introduced recently by Feng and Li~[ICALP 2024] for the additive valuation case. Our rounding algorithm is similar to that of Li~[SODA 2025] developed for the unrelated machine scheduling problem to minimize weighted completion time. Roughly speaking, we designate the largest item in each configuration as a large item and the remaining items as small items. So, every agent gets precisely 1 fractional large item in the configuration LP solution. With the rounding algorithm in Li~[SODA 2025], we can ensure that in the obtained solution, every agent gets precisely 1 large item, and the assignments of small items are negatively correlated.

cs.GT↗

arXiv2025

MoEntwine: Unleashing the Potential of Wafer-scale Chips for Large-scale Expert Parallel Inference

As large language models (LLMs) continue to scale up, mixture-of-experts (MoE) has become a common technology in SOTA models. MoE models rely on expert parallelism (EP) to alleviate memory bottleneck, which introduces all-to-all communication to dispatch and combine tokens across devices. However, in widely-adopted GPU clusters, high-overhead cross-node communication makes all-to-all expensive, hindering the adoption of EP. Recently, wafer-scale chips (WSCs) have emerged as a platform integrating numerous devices on a wafer-sized interposer. WSCs provide a unified high-performance network connecting all devices, presenting a promising potential for hosting MoE models. Yet, their network is restricted to a mesh topology, causing imbalanced communication pressure and performance loss. Moreover, the lack of on-wafer disk leads to high-overhead expert migration on the critical path. To fully unleash this potential, we first propose Entwined Ring Mapping (ER-Mapping), which co-designs the mapping of attention and MoE layers to balance communication pressure and achieve better performance. We find that under ER-Mapping, the distribution of cold and hot links in the attention and MoE layers is complementary. Therefore, to hide the migration overhead, we propose the Non-invasive Balancer (NI-Balancer), which splits a complete expert migration into multiple steps and alternately utilizes the cold links of both layers. Evaluation shows ER-Mapping achieves communication reduction up to 62%. NI-Balancer further delivers 54% and 22% improvements in MoE computation and communication, respectively. Compared with the SOTA NVL72 supernode, the WSC platform delivers an average 39% higher per-device MoE performance owing to its scalability to larger EP.

cs.DC↗

arXiv2025

ComDrive: Comfort-Oriented End-to-End Autonomous Driving

We propose ComDrive: the first comfort-oriented end-to-end autonomous driving system to generate temporally consistent and comfortable trajectories. Recent studies have demonstrated that imitation learning-based planners and learning-based trajectory scorers can effectively generate and select safety trajectories that closely mimic expert demonstrations. However, such trajectory planners and scorers face the challenge of generating temporally inconsistent and uncomfortable trajectories. To address these issues, ComDrive first extracts 3D spatial representations through sparse perception, which then serves as conditional inputs. These inputs are used by a Conditional Denoising Diffusion Probabilistic Model (DDPM)-based motion planner to generate temporally consistent multi-modal trajectories. A dual-stream adaptive trajectory scorer subsequently selects the most comfortable trajectory from these candidates to control the vehicle. Experiments demonstrate that ComDrive achieves state-of-the-art performance in both comfort and safety, outperforming UniAD by 17% in driving comfort and reducing collision rates by 25% compared to SparseDrive. More results are available on our project page: https://jmwang0117.github.io/ComDrive/.

cs.CV↗

arXiv2025

Nearly Optimal Bounds for Stochastic Online Sorting

In the online sorting problem, we have an array $A$ of $n$ cells, and receive a stream of $n$ items $x_1,\dots,x_n\in [0,1]$. When an item arrives, we need to immediately and irrevocably place it into an empty cell. The goal is to minimize the sum of absolute differences between adjacent items, which is called the \emph{cost} of the algorithm. It has been shown by Aamand, Abrahamsen, Beretta, and Kleist (SODA 2023) that when the stream $x_1,\dots,x_n$ is generated adversarially, the optimal cost bound for any deterministic algorithm is $Θ(\sqrt{n})$. In this paper, we study the stochastic version of online sorting, where the input items $x_1,\dots,x_n$ are sampled uniformly at random. Despite the intuition that the stochastic version should yield much better cost bounds, the previous best algorithm for stochastic online sorting by Abrahamsen, Bercea, Beretta, Klausen and Kozma (ESA 2024) only achieves $\tilde{O}(n^{1/4})$ cost, which seems far from optimal. We show that stochastic online sorting indeed allows for much more efficient algorithms, by presenting an algorithm that achieves expected cost $\log n\cdot 2^{O(\log^* n)}$. We also prove a cost lower bound of $Ω(\log n)$, thus show that our algorithm is nearly optimal.

cs.DS↗

arXiv2025

Static Retrieval Revisited: To Optimality and Beyond

In the static retrieval problem, a data structure must answer retrieval queries mapping a set of $n$ keys in a universe $[U]$ to $v$-bit values. Information-theoretically, retrieval data structures can use as little as $nv$ bits of space. For small value sizes $v$, it is possible to achieve $O(1)$ query time while using space $nv + o(n)$ bits -- whether or not such a result is possible for larger values of $v$ (e.g., $v = Θ(\log n)$) has remained open. In this paper, we obtain a tight lower bound (as well as matching upper bounds) for the static retrieval problem. In the case where values are large, we show that there is actually a significant tension between time and space. It is not possible, for example, to get $O(1)$ query time using $nv + o(n)$ bits of space, when $v = Θ(\log n)$ (and assuming the word RAM model with $O(\log n)$-bit words). At first glance, our lower bound would seem to render retrieval unusable in many settings that aim to achieve very low redundancy. However, our second result offers a way around this: We show that, whenever a retrieval data structure $D_1$ is stored along with another data structure $D_2$ (whose size is similar to or larger than the size of $D_1$), it is possible to implement the combined data structure $D_1 \cup D_2$ so that queries to $D_1$ take $O(1)$ time, operations on $D_2$ take the same asymptotic time as if $D_2$ were stored on its own, and the total space is $nv + \mathrm{Space}(D_2) + n^{0.67}$ bits.

cs.DS↗

arXiv2025

From Quarter to All: Accelerating Speculative LLM Decoding via Floating-Point Exponent Remapping and Parameter Sharing

Large language models achieve impressive performance across diverse tasks but exhibit high inference latency due to their large parameter sizes. While quantization reduces model size, it often leads to performance degradation compared to the full model. Speculative decoding remains lossless but typically incurs extra overheads. We propose SPEQ, an algorithm-hardware co-designed speculative decoding method that uses part of the full-model weight bits to form a quantized draft model, thereby eliminating additional training or storage overhead. A reconfigurable processing element array enables efficient execution of both the draft and verification passes. Experimental results across 15 LLMs and tasks demonstrate that SPEQ achieves speedups of 2.07x, 1.53x, and 1.45x compared over FP16, Olive, and Tender, respectively.

cs.AR↗