arXiv subjects

Xiang Hu

Publications and source records attributed to Xiang Hu.

At least 19 recordsLinked to original sources

arXiv2026

DA-DLM: Explicitly Modeling Token Dependencies in Diffusion Language Models

Diffusion Language Models (DLMs) generate text by iteratively denoising a masked sequence, independently predicting multiple tokens at each step. This conditional independence discards inter-token dependencies and degrades coherence-an issue that parallels the multi-modality problem in Non-Autoregressive Translation (NAT). Drawing on the Directed Acyclic Transformer (DAT), which tackles this problem in NAT via a Directed Acyclic Graph (DAG), we propose DA-DLM, a model that adapts DAG-based dependency modeling to DLMs' iterative setting through a position-oriented DAG design. The position-oriented DAG binds node groups to fixed output positions so that tokens fixed in earlier steps anchor neighboring predictions via learned transitions, and evolves with denoising to focus on remaining uncertainty as anchors accumulate. On language modeling, open-ended generation, and summarization, DA-DLM consistently outperforms Block Diffusion, especially under fewer denoising steps, and matches autoregressive models while preserving the parallel generation advantage. Our code is publicly available at https://github.com/jipy0222/DA-DLM.

cs.CL↗

arXiv2026

SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking

Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (tokens or blocks) for each query. Existing trainable methods usually use a lightweight selector to score context units, followed by hard Top-K selection that blocks gradients from the language modeling loss. Consequently, these methods commonly distill layer-wise dense attention distributions. Although this encourages the selector to rank context units by dense attention weights in the original model, the ranking is not directly aligned with their impact on predictions under a fixed attention budget (i.e., the number of attended context units per query), potentially wasting the limited budget on less useful units. To address this misalignment, we propose Simple Attention Sparsification (SAS), a gated sparse attention mechanism that optimizes context ranking end-to-end with the language modeling loss. The key idea is to inject the selector's continuous scores into attention logits during training, allowing the loss to update the selector through standard backpropagation. We identify several choices crucial for this simple design to work well in practice: placing the gate inside the attention softmax in log form, using normalized softmax gates to calibrate historical context against the always-retained current block, and preserving continuous selector scores so the model learns relative priorities rather than only hard selections. To support long-sequence training, we implement a memory-efficient Triton kernel that integrates SAS into FlashAttention-style computation. Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

cs.CL↗

arXiv2026

FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose \textbf{Lookahead Sparse Attention (LSA)}, a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a \textbf{backbone-free decoupled training} strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this ``less is more'' paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), \texttt{FM-DS-V4} compresses the average physical KV cache footprint down to merely 13.5\% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6\% absolute margin on average). At 1M context, per-decode-token compute drops to 0.30$\times$ of the baseline and GPU KV cache shrinks by 90\% (3.73$\to$0.37 GB), translating into \textbf{2.8$\times$ aggregate throughput and 2.7$\times$ concurrency gains} in PD-disaggregated serving on 8$\times$H20 GPUs.

cs.LG↗

arXiv2026

Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

cs.CL↗

arXiv2026

SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification

Aerial-Ground Person Re-IDentification (AG-ReID) aims to retrieve specific persons across cameras with different viewpoints. Previous works focus on designing discriminative models to maintain the identity consistency despite drastic changes in camera viewpoints. The core idea behind these methods is quite natural, but designing a view-robust model is a very challenging task. Moreover, they overlook the contribution of view-specific features in enhancing the model's ability to represent persons. To address these issues, we propose a novel generative framework named SD-ReID for AG-ReID, which leverages generative models to mimic the feature distribution of different views while extracting robust identity representations. More specifically, we first train a ViT-based model to extract person representations along with controllable conditions, including identity and view conditions. We then fine-tune the Stable Diffusion (SD) model to enhance person representations guided by these controllable conditions. Furthermore, we introduce the View-Refined Decoder (VRD) to bridge the gap between instance-level and global-level features. Finally, both person representations and all-view features are employed to retrieve target persons. Extensive experiments on five AG-ReID benchmarks (i.e., CARGO, AG-ReIDv1, AG-ReIDv2, LAGPeR and G2APS-ReID) demonstrate the effectiveness of our proposed method. The source code and pre-trained models are available at https://github.com/924973292/SD-ReID.

cs.CV↗

arXiv2026

Understanding and Improving Length Generalization in Hierarchical Sparse Attention Models

Effectively processing long contexts is a critical challenge for language models. While standard Transformers are limited by quadratic complexity and poor length extrapolation, alternative architectures like sliding window attention and state space models sacrifice the ability to effectively utilize the full context due to their fixed-size memory. Chunk-based sparse attention has emerged as a promising paradigm for extreme length generalization, yet the key architectural principles underpinning its success are not yet fully understood. In this work, we present a systematic dissection of these models to identify the core components driving their performance. Through a unified framework and comprehensive ablation studies, we demonstrate that a combination of three design principles is critical: (1) an expressive, non-linear Chunk Encoder with a dedicated CLS token to produce representations for retrieval; (2) a Bypassing Residual Path to stably integrate retrieved global information without it being overridden by the local residual stream; and (3) enforced selection sparsity during pre-training to bridge the train-test distribution gap. We provide a theoretical motivation for intra-chunk information processing and landmark generation. By combining these principles, we establish a new state-of-the-art for training-free length extrapolation, successfully generalizing models trained on a 4K context to 32 million tokens on RULER and BABILong. Our findings provide a clear and empirically-grounded set of design principles for developing future, highly-capable long-context language models.

cs.CL↗

arXiv2025

aiXiv: A Next-Generation Open Access Ecosystem for Scientific Discovery Generated by AI Scientists

Recent advances in large language models (LLMs) have enabled AI agents to autonomously generate scientific proposals, conduct experiments, author papers, and perform peer reviews. Yet this flood of AI-generated research content collides with a fragmented and largely closed publication ecosystem. Traditional journals and conferences rely on human peer review, making them difficult to scale and often reluctant to accept AI-generated research content; existing preprint servers (e.g. arXiv) lack rigorous quality-control mechanisms. Consequently, a significant amount of high-quality AI-generated research lacks appropriate venues for dissemination, hindering its potential to advance scientific progress. To address these challenges, we introduce aiXiv, a next-generation open-access platform for human and AI scientists. Its multi-agent architecture allows research proposals and papers to be submitted, reviewed, and iteratively refined by both human and AI scientists. It also provides API and MCP interfaces that enable seamless integration of heterogeneous human and AI scientists, creating a scalable and extensible ecosystem for autonomous scientific discovery. Through extensive experiments, we demonstrate that aiXiv is a reliable and robust platform that significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing on aiXiv. Our work lays the groundwork for a next-generation open-access ecosystem for AI scientists, accelerating the publication and dissemination of high-quality AI-generated research content. Code: https://github.com/aixiv-org aiXiv: https://aixiv.science

cs.AI↗

arXiv2025

Flash Multi-Head Feed-Forward Network

We explore Multi-Head FFN (MH-FFN) as a replacement of FFN in the Transformer architecture, motivated by the structural similarity between single-head attention and FFN. While multi-head mechanisms enhance expressivity in attention, naively applying them to FFNs faces two challenges: memory consumption scaling with the head count, and an imbalanced ratio between the growing intermediate size and the fixed head dimension as models scale, which degrades scalability and expressive power. To address these challenges, we propose Flash Multi-Head FFN (FlashMHF), with two key innovations: an I/O-aware fused kernel computing outputs online in SRAM akin to FlashAttention, and a design using dynamically weighted parallel sub-networks to maintain a balanced ratio between intermediate and head dimensions. Validated on models from 128M to 1.3B parameters, FlashMHF consistently improves perplexity and downstream task accuracy over SwiGLU FFNs, while reducing peak memory usage by 3-5x and accelerating inference by up to 1.08x. Our work establishes the multi-head design as a superior architectural principle for FFNs, presenting FlashMHF as a powerful, efficient, and scalable alternative to FFNs in Transformers.

cs.LG↗

arXiv2025

Every Token Counts: Generalizing 16M Ultra-Long Context in Large Language Models

This work explores the challenge of building ``Machines that Can Remember'', framing long-term memory as the problem of efficient ultra-long context modeling. We argue that this requires three key properties: \textbf{sparsity}, \textbf{random-access flexibility}, and \textbf{length generalization}. To address ultra-long-context modeling, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into Transformers to build HSA-UltraLong, which is an 8B-parameter MoE model trained on over 8 trillion tokens and is rigorously evaluated on different tasks with in-domain and out-of-domain context lengths to demonstrate its capability in handling ultra-long contexts. Results show that our model performs comparably to full-attention baselines on in-domain lengths while achieving over 90\% accuracy on most in-context retrieval tasks with contexts up to 16M. This report outlines our experimental insights and open problems, contributing a foundation for future research in ultra-long context modeling.

cs.CL↗

arXiv2025

Hardware-aligned Hierarchical Sparse Attention for Efficient Long-term Memory Access

A key advantage of Recurrent Neural Networks (RNNs) over Transformers is their linear computational and space complexity enables faster training and inference for long sequences. However, RNNs are fundamentally unable to randomly access historical context, and simply integrating attention mechanisms may undermine their efficiency advantages. To overcome this limitation, we propose Hierarchical Sparse Attention (HSA), a novel attention mechanism that enhances RNNs with long-range random access flexibility while preserving their merits in efficiency and length generalization. HSA divides inputs into chunks, selects the top-$k$ chunks and hierarchically aggregates information. The core innovation lies in learning token-to-chunk relevance based on fine-grained token-level information inside each chunk. This approach enhances the precision of chunk selection across both in-domain and out-of-domain context lengths. To make HSA efficient, we further introduce a hardware-aligned kernel design. By combining HSA with Mamba, we introduce RAMba, which achieves perfect accuracy in passkey retrieval across 64 million contexts despite pre-training on only 4K-length contexts, and significant improvements on various downstream tasks, with nearly constant memory footprint. These results show RAMba's huge potential in long-context modeling.

cs.CL↗

arXiv2025

LATex: Leveraging Attribute-based Text Knowledge for Aerial-Ground Person Re-Identification

As an important task in intelligent transportation systems, Aerial-Ground person Re-IDentification (AG-ReID) aims to retrieve specific persons across heterogeneous cameras in different viewpoints. Previous methods typically adopt deep learning-based models, focusing on extracting view-invariant features. However, they usually overlook the semantic information in person attributes. In addition, existing training strategies often rely on full fine-tuning large-scale models, which significantly increases training costs. To address these issues, we propose a novel framework named LATex for AG-ReID, which adopts prompt-tuning strategies to leverage attribute-based text knowledge. Specifically, with the Contrastive Language-Image Pre-training (CLIP) model, we first propose an Attribute-aware Image Encoder (AIE) to extract both global semantic features and attribute-aware features from input images. Then, with these features, we propose a Prompted Attribute Classifier Group (PACG) to predict person attributes and obtain attribute representations. Finally, we design a Coupled Prompt Template (CPT) to transform attribute representations and view information into structured sentences. These sentences are processed by the text encoder of CLIP to generate more discriminative features. As a result, our framework can fully leverage attribute-based text knowledge to improve AG-ReID performance. Extensive experiments on three AG-ReID benchmarks demonstrate the effectiveness of our proposed methods. The source code is available at https://github.com/kevinhu314/LATex.

cs.CV↗

arXiv2025

Laser-Induced Commensurate-Incommensurate Transition of Charge Order in a Hubbard Superlattice

We investigate the nonequilibrium dynamics of charge density waves in a pumped one-dimensional Hubbard superlattice with staggered onsite Coulomb interactions at half-filling, using time-dependent exact diagonalization. In equilibrium, the system exhibits commensurate charge correlations consistent with the superlattice periodicity. Under laser excitation, the charge correlation function exhibits distinct behaviors across four representative frequencies, spanning both linear and nonlinear optical regimes. Notably, we observe a laser-induced commensurate-to-incommensurate transition in the charge order, manifested by a shift in the peak wavevector of the charge structure factor. This transition is driven by sublattice-selective doublon-holon dynamics, where the laser frequency and intensity determine whether excitations predominantly destabilize the charge order on the weakly or strongly interacting sublattice. Our analysis of the excitation spectrum and site-resolved correlation dynamics reveals the underlying mechanisms of this transition. These results suggest a promising optical strategy for controlling charge order in superlattice-based quantum materials.

cond-mat.str-el↗

arXiv2025

Effect of Inversion Asymmetry on Bilayer Graphene's Superconducting and Exciton Condensates

Inversion asymmetry in bilayer graphene can be tuned by the displacement field. As a result, the band dispersion in biased bilayer graphene acquires flat band regions near the Dirac points along with a non-trivial band geometry. We analyze the effect of inversion symmetry on the critical temperature and superfluid stiffness of the superconducting state of AB-stacked graphene bilayer and on the exciton condensate in double layers formed by two AB-stacked graphene bilayers. The geometric superfluid stiffness in bilayer graphene superconductors is found to be negligible due to the small superconducting gap. Furthermore, we show that the geometric superfluid stiffness is maximized for a constant order parameter. Therefore, it can be neglected in biased bilayer graphene superconductors with any pairing symmetry. However, the displacement field enhances the geometric superfluid stiffness in exciton condensates. It is most prominent at low densities and high displacement fields. A consequence of the geometric superfluid stiffness is a modest enhancement of the Berezinskii-Kosterlitz-Thouless transition temperature in bilayer graphene's exciton condensate.

cond-mat.mes-hall↗

arXiv2025

Unsupervised Morphological Tree Tokenizer

As a cornerstone in language modeling, tokenization involves segmenting text inputs into pre-defined atomic units. Conventional statistical tokenizers often disrupt constituent boundaries within words, thereby corrupting semantic information. To address this drawback, we introduce morphological structure guidance to tokenization and propose a deep model to induce character-level structures of words. Specifically, the deep model jointly encodes internal structures and representations of words with a mechanism named $\textit{MorphOverriding}$ to ensure the indecomposability of morphemes. By training the model with self-supervised objectives, our method is capable of inducing character-level structures that align with morphological rules without annotated training data. Based on the induced structures, our algorithm tokenizes words through vocabulary matching in a top-down manner. Empirical results indicate that the proposed method effectively retains complete morphemes and outperforms widely adopted methods such as BPE and WordPiece on both morphological segmentation tasks and language modeling tasks. Code is available at https://github.com/martianmartina/TreeTokenizer.

cs.CL↗

arXiv2025

A Systematic Study of Compositional Syntactic Transformer Language Models

Syntactic language models (SLMs) enhance Transformers by incorporating syntactic biases through the modeling of linearized syntactic parse trees alongside surface sentences. This paper focuses on compositional SLMs that are based on constituency parse trees and contain explicit bottom-up composition of constituent representations. We identify key aspects of design choices in existing compositional SLMs and propose a unified framework encompassing both existing models and novel variants. We conduct a comprehensive empirical evaluation of all the variants in our framework across language modeling, syntactic generalization, summarization, dialogue, and inference efficiency. Based on the experimental results, we make multiple recommendations on the design of compositional SLMs. Our code is released at https://github.com/zhaoyd1/compositional_SLMs.

cs.CL↗

arXiv2025

Efficient Length-Generalizable Attention via Causal Retrieval for Long-Context Language Modeling

Despite the success of Transformers, handling long contexts remains challenging due to the limited length generalization and quadratic complexity of self-attention. Thus Transformers often require post-training with a larger attention window, significantly increasing computational and memory costs. In this paper, we propose a novel attention mechanism based on dynamic context, Grouped Cross Attention (GCA), which can generalize to 1000 times the pre-training context length while maintaining the ability to access distant information with a constant attention window size. For a given input sequence, we split it into chunks and use each chunk to retrieve top-k relevant past chunks for subsequent text generation. Specifically, unlike most previous works that use an off-the-shelf retriever, our key innovation allows the retriever to learn how to retrieve past chunks that better minimize the auto-regressive loss of subsequent tokens in an end-to-end manner. Such a mechanism accommodates retrieved chunks with a fixed-size attention window to achieve long-range information access, significantly reducing computational and memory costs during training and inference. Experiments show that GCA-based models achieve near-perfect accuracy in passkey retrieval for 16M context lengths, which is 1000 times the training length.

cs.CL↗

arXiv2025

Analysis of Radiation Level and Estimation of Protection Distance of γ Mobile Flaw Detection Source

Objective To analyze the radiation dose associated with gamma-ray mobile flaw detection, estimate the extent of the supervision and control areas, and assess the associated radiation risks. Methods A combination of theoretical calculations and actual measurements was used to compare and analyze the ambient equivalent dose rates of 192 Ir and 75 Se at their nominal source strengths. Measurements were conducted at distances of 1 m, 2 m, and 5 m from the radiation source. The extents of the control and supervision areas were estimated under three working scenarios: 1 without considering air attenuation, 2 considering air attenuation, and 3 after shielding by the flaw detection workpiece, using source activities of 3.7 * 10^10 Bq and 3.7 * 10^12Bq. Results Actual measurement of radiation dose of 192 Ir and 75 Se were measured under three different nominal activities. Theoretical calculation of radiation dose estimates at various distances were obtained for both nuclides, and the results showed that the theoretical values were basically consistent with the measured values. Conclusion The estimated scope of the supervision and control areas provided in this study can serve as a reference for flaw detection companies. Technicians can use these estimates to calculate appropriate distances for safety zones based on different nuclide activities. This enables flaw detection personnel to reduce the measurement scope on-site and to quickly and accurately define area boundaries.

physics.ins-det↗

arXiv2024

A new pathway to impact ionization in a photo-excited one-dimensional ionic Hubbard model

Using the time-dependent Lanczos method, we study the non-equilibrium dynamics of the half-filled one-dimensional ionic Hubbard model, deep within the Mott insulating regime, under the influence of a transient laser pulse. In equilibrium, increasing the staggered potential in the Mott regime reduces the Mott gap and broadens the Hubbard bands, creating favorable conditions for impact ionization. After laser excitation, impact ionization is observed, with its occurrence depending on both the staggered potential and the laser pump frequency. By analyzing the time evolution of the kinetic, ionic, and Coulomb interaction energies, we identify a novel mechanism for impact ionization, in which excess ionic potential energy is converted into additional double occupancy-distinct from the conventional mechanism where excess kinetic energy drives this process. We further show that impact ionization arises from interference between excited states driven by photon excitation of the same order. These results present a new pathway for realizing impact ionization in strongly correlated electron systems.

cond-mat.str-el↗