Search arXiv⌕ Search

arXiv subjects

Yu Cheng

Publications and source records attributed to Yu Cheng.

At least 55 records · Page 3Linked to original sources
arXiv2026

DeepStack: Facilitating Co-Design Exploration of 3D DRAM-Stacked Accelerators for Distributed LLM Inference

Advances in hybrid bonding and packaging have driven growing interest in 3D DRAM-stacked AI accelerators. As large language models (LLMs) scale to hundreds of billions or trillions of parameters, distributed inference across multiple 3D chips has become essential for AI serving. This trend makes cross-stack co-design critical because system-level parallelization and scheduling choices are tightly coupled with hardware characteristics such as memory organization, interconnects, and thermal constraints. We present DeepStack, an accurate performance model and efficient design space exploration (DSE) framework for distributed 3D-stacked LLM inference. At the hardware level, DeepStack captures transaction-aware memory bandwidth, bank activation constraints, buffering limitations, and thermal and power behavior. At the system level, it incorporates comprehensive parallelization strategies and execution scheduling. Through a dual-stage network abstraction and tile-level compute-communication overlap modeling, DeepStack achieves up to 100,000x faster evaluation than state-of-the-art simulators at comparable accuracy. We cross-validate DeepStack against our in-house 3D designs, an NS-3 backend with 2.12% error, and vLLM serving on eight B200 GPUs with 12.92% error. Combined with hierarchical search, DeepStack efficiently explores about 2.5 x 10^14 design points spanning the number of stacked DRAM layers, DRAM vertical connectivity, interconnects, compute-memory allocation, and distributed scheduling under thermal and area constraints. A search-space ablation shows that restricted DSE baselines can miss up to 9.5x modeled throughput. Beyond modeling and DSE, DeepStack derives design implications for distributed 3D AI systems and guides performance optimization across the stack. Source code and artifacts are available at https://github.com/tile-ai/DeepStack/tree/ae.

cs.AR↗
arXiv2026

Critical-Window Fluctuations and Disorder Universality for the Sherrington--Kirkpatrick Model

We establish free-energy fluctuation limits for the Ising Sherrington--Kirkpatrick model in the nonzero parts of its critical window. For fixed $b\ne0$ and $β_N=1+bN^{-1/3}\sqrt{\log N}$, our main Gaussian orthogonal ensemble (GOE) result is \[ \sqrt{\frac6{\log N}}\left(F_{N,β_N}-N\,\mathrm{FE}(β_N)+\frac{\log N}{12}\right)\xrightarrow{d}G+\sqrt{\frac32}\,b_+TW_1, \] where $G$ is standard Gaussian, $TW_1$ has the real Tracy--Widom law, and $G$ is independent of $TW_1$, and $\mathrm{FE}$ denotes the limit of spherical Sherrington--Kirkpatrick free-energy. Additionally, we show that in a moderately supercritical regime \[ \frac{2}{N^{1/3}(β_N-1)}\left(F_{N,β_N}-N\,\mathrm{FE}(β_N)+\frac{\log N}{12}\right)\xrightarrow{d}TW_1. \] We also show that the above results remains valid for independent, not necessarily identically distributed, disorder matrices whose first three moments match the Gaussian law and whose fourth moments satisfy an averaged bound.

math.PR↗
arXiv2026

An Eccentric Massive Protobinary Assembled via a Core-merger Parabolic Encounter

Most massive stars form in binary systems, which profoundly influence their subsequent evolution. However, how such systems form remains poorly understood, with several competing scenarios proposed, including disk fragmentation, core fragmentation and capture. Determining the orbital architectures of massive binaries, particularly during their earliest embedded phases, is therefore crucial for distinguishing among these formation pathways, but direct measurements of their three-dimensional motions have remained exceptionally challenging. Here we present high-resolution, multi-epoch sub-millimeter-to-centimeter ALMA and JVLA observations of the massive protobinary IRAS 07299$-$1651, complemented by JWST and VLT infrared imaging. We detect orbital proper motion of the binary components, enabling a full three-dimensional orbital reconstruction. Combining orbital fitting, multi-wavelength continuum modelling, hydrogen recombination line kinematics and jet observations, we find that the preferred orbital solutions are highly eccentric and close to parabolic, while both circumstellar disks are strongly misaligned with the orbital plane. These properties are naturally explained by a ``core-merger'' scenario in which the two protostars originated independently from initially unbound cores that recently underwent a near-parabolic encounter, producing an eccentric binary with a current separation of about 200 au. These findings suggest that the core-merger process may represent an important pathway for forming eccentric massive binaries.

astro-ph.SR↗
arXiv2026

Magnetic Fields in Massive Star-forming Regions (MagMaR). IX. Radiative Torque Alignment and Disruption in NGC6334I

Intense radiation from high-mass stars is expected to significantly affect dust grain alignment and evolution through RAdiative Torques (RATs). We investigate this effect in a massive star-forming region, NGC6334I, using 1.2 mm dust continuum polarization observations from the Atacama Large Millimeter/submillimeter Array. The polarization fraction spans from $\lesssim1\%$ to $\sim10\%$ and decreases with increasing column density, remaining below $2\%$ in dense cores despite high temperatures ($\sim100$ K), where efficient grain alignment by RATs is expected. We investigate how grain alignment, grain growth, grain disruption, B-field tangling, and local physical conditions affect the polarization properties of MM1, MM2, MM3, and their surroundings. Polarization angle dispersion shows that B-field tangling contributes to depolarization at moderate densities but cannot fully explain the lowest polarization fractions. Using RAT-based grain alignment and polarization modeling, we find that reduced alignment efficiency and high optical depth reproduce the low polarization in the densest regions. MM2 shows evidence of grain growth, with maximum grain sizes $a_{\max}\sim0.35-1.0~μ$m, while MM1 exhibits smaller values of $\sim0.35-0.50~μ$m. Accounting for optical depth increases the inferred grain sizes in MM1 to $\sim1.0-2.0~μ$m. Analytical estimates of radiative torque disruption from the intense outburst suggest that micron-sized grains in high-temperature, moderate-density regions can fragment into submicron grains. Alternatively, high optical depth may also explain the low polarization in the densest regions even in the presence of micron-sized grains. Incorporating the B-field inclination effect indicates a transition from predominantly plane-of-sky fields at low densities to more line-of-sight-aligned configurations at high densities.

astro-ph.GA↗
arXiv2026

Cross-Task Generalization Between Understanding and Generation in Unified Vision-Language Models: A Controlled Study

Unified vision-language models (VLMs) aim to support both visual understanding and generation within a single framework, but it remains unclear when mixed training benefits both capabilities and when it introduces conflicts. This paper presents a controlled empirical study of cross-task generalization between understanding and generation in unified VLMs. We construct two controllable image-text benchmarks, SmartWatch and modified CelebA, with paired VQA, captioning, and text-to-image generation tasks, and evaluate multiple LLM-based unified architectures built from SigLIP and VQ-VAE visual spaces. Our experiments show that mixed understanding-generation training can improve both tasks over task-specific training, but the benefit depends strongly on the relation between vision input and output spaces. Unified models with better aligned visual spaces exhibit stronger cross-task transfer, while reversible affine distortions of the input visual space substantially weaken this effect and can turn mutual benefits into conflicts. We further find that increasing data from one task can initially improve the other, but excessive imbalance between understanding and generation data may degrade the complementary task. By controlling attribute frequencies, we show that generation supervision can help recover underrepresented visual concepts for understanding. Adapter analyses suggest that this transfer is not primarily caused by richer visual adapter features, but by the base language model learning relationships that generalize across aligned visual spaces. A real-case experiment on LLaVA provides additional evidence that mixed understanding-generation training can benefit visual understanding beyond controlled benchmarks.

cs.CV↗
arXiv2026

Higher-Dimensional Rotary Position Embedding

Transformers rely on position embedding mechanisms in long context modeling in most cases. Rotary Position Embedding (RoPE) embeds positional information with independent 2D rotations, forming relative position terms in self-attention. However, its pairwise, block-based, and decoupled structure limits deep mixing and robustness across channels. We propose HD-RoPE, which extends RoPE from independent 2D rotations to higher-dimensional rotations and introduces a Paley-I orthogonal basis to obtain balanced, isotropic, and dense phase mixing within each rotation subspace. This significantly enhances channel coupling and rotational degrees of freedom while maintaining orthogonal stability and the relative position closure property. Furthermore, HD-RoPE is easily optimized for engineering efficiency without introducing additional trainable parameters. We have conducted extensive evaluation results demonstrating that HD-RoPE achieves significant performance improvements over standard RoPE across various popular benchmarks and in both long and short contexts.

cs.LG↗
arXiv2026

Think, Look, and Revise: Inconsistency-Aware Visual Self-Correction in MLLMs

Tool-augmented multimodal reasoning integrates external tools (e.g., object detection, depth estimation) into multimodal large language models (MLLMs) to address perceptual bottlenecks in complex visual tasks. However, existing approaches rarely verify tool outputs, limiting their ability to detect and recover from tool failures. We propose ReVISE, a framework that equips MLLMs with verification and dynamic error recovery for tool-augmented reasoning. ReVISE introduces (1) a curated training dataset that supervises reflective behaviors, enabling models to validate tool-derived evidence, reformulate queries when visual mismatches arise, and fall back to intrinsic grounding when external tools are unreliable; and (2) a reinforcement learning based targeted rewards that encourage internal reflection and penalize spatial misalignment. Experiments on several benchmarks demonstrate consistent improvements over existing methods, highlighting the importance of error detection and correction in tool-augmented multimodal reasoning.

cs.CV↗
arXiv2026

Dense Cores in the Vicinity of an HII Region

Massive stars strongly influence their surroundings through radiative and mechanical feedback, but its effects on dense gas structures at sub-pc scales remain poorly constrained. We investigate how feedback from a newly formed massive star affects dense cores in the filamentary molecular cloud IRAS 18530+0215. We analyze ALMA Band 6 observations of 1.3 mm dust continuum and DCN, N$_2$D$^+$, and $^{13}$CS line emission, together with VLA K-band continuum and NH$_3$ observations. Dense cores are identified with astrodendro, and their temperatures, masses, velocity dispersions, and virial parameters are derived. The dynamical state of the ultra-compact H II region is examined through energy and pressure estimates. The H II region has a radius of $\sim$0.1 pc and an expansion velocity of $\sim$2.5 km s$^{-1}$, corresponding to a shell dynamical age of $\sim$0.06 Myr. DCN and $^{13}$CS cores are concentrated near the H II region, whereas N$_2$D$^+$ cores preferentially lie farther away. Core temperatures and velocity dispersions decrease with projected distance from the H II region. Virial parameters increase within the inner $\sim$0.3 pc but decline sharply beyond this scale, while core masses show no significant trend with distance. Strong star formation signatures are found at $\sim$0.2 pc, whereas more distant regions still host quiescent, cold dense cores. The compact H II region appears trapped or choked within $\sim$0.1 pc, while its feedback extends to at least $\sim$0.3 pc. Within this region, feedback enhances core velocity dispersions, gas temperatures, and virial parameters, with no evidence that it promotes the formation of more massive dense cores.

astro-ph.GA↗
arXiv2026

CoinVE-200K: A Large-Scale High-Quality Dataset for Compositional Instruction-Guided Video Editing

The quality and diversity of instruction-based video editing datasets are steadily improving, yet existing datasets mainly focus on single editing operations and fall short in supporting compositional instruction-guided video editing. In particular, multiple editing intents must be jointly understood and faithfully executed within the same video. To address this issue, we introduce CoinVE-200K, a large-scale, high-quality dataset for Compositional Instruction-Guided Video Editing. CoinVE-200K contains 1080p video-editing pairs of up to 201 frames, covering diverse compositional scenarios where each sample involves 2 to 5 atomic editing operations. The instructions target humans, objects, and backgrounds, and cover edit types such as addition, removal, modification, and stylization. All samples are built through a carefully designed generation and filtering pipeline to ensure instruction faithfulness, visual quality, temporal consistency, and compositional diversity. We also introduce CoinVE-Bench, a benchmark for compositional-instruction video editing across diverse subjects, operation types, and instruction complexities. Furthermore, we present CoinVE-Edit, a 22B compositional video editing model built upon Wan2.1-T2V-14B and Qwen3-VL-8B-Instruct. CoinVE-Edit disentangles region-aware attention for different editing instructions, enabling precise multi-region editing while preserving irrelevant content and temporal coherence. Experiments on CoinVE-Bench show that CoinVE-Edit achieves strong performance in instruction following, compositional editing accuracy, visual quality, and temporal consistency.

cs.CV↗
arXiv2026

Towards a Densing Law for User Representation Learning at Billion-Scale Capacity

User representation learning in real-world industrial scenarios is commonly scaled by increasing user amount, behavioral sequence length and model size. However, existing methods face two challenges: (i) Bottleneck for raw data scaling at billion-scale capacity, as performance exhibit diminishing performance gains with larger-scale raw text user behavioral input, which can be mitigated by tokenization. (ii) Lack of quantitative analysis of how tokenization configurations should scale with data size. In this report, we propose User Behavioral Densing Law for characterizing the quantitative relationship between data scale and the minimum sufficient tokenization capacity. Firstly, we conduct a pilot study on raw & tokenized scaling comparison on billion-scale Alipay dataset, revealing the raw data scaling bottleneck and the sustained gains enabled by tokenization. To derive the scaling pattern governing the minimum sufficient tokenization configuration at different data scales, theoretical analysis and systematic experiments are employed to summarize the quantitative scaling pattern. We find an approximately linear relationship between the logarithms of minimum sufficient tokenization capacity and input data size measured by tokens, and the scaling slope varies systematically with the tokenization method and data source, reflecting differences in representation-space redundancy and intra-source uniqueness. Guided by the proposed law, we further develop ALGN, an adaptive variable-length tokenization method that improves capacity allocation. Extensive experiments across diverse data sources, tokenization methods, and downstream tasks demonstrate the generalizability and reliability of the User Behavioral Densing Law, providing practical guidance for tokenization configuration selection in large-scale user representation learning. Moreover, ALGN outperforms existing baselines.

cs.IR↗
arXiv2026

The dominance of turbulence over magnetism in the formation of massive star cluster seeds

High-mass stars form in protoclusters, where gravo-magnetic processes shape collapsing clouds and clumps to be elongated preferentially perpendicular to magnetic (B) fields. Yet it remains unclear whether gravo-magnetic processes still govern the formation of smaller-scale condensations in massive-star-forming protoclusters, which are crucial for understanding the stellar initial mass function and multiplicity. Here we report the first statistical evidence that the condensation elongations are preferentially aligned with local B fields, based on high-resolution data from the largest dust polarization survey toward 30 massive star-forming regions with the Atacama Large Millimeter/submillimeter Array (ALMA). Our clustered massive star formation simulations reveal that this more parallel alignment is exclusively observed in models where initial turbulence dominates B fields. In contrast, models with initial B fields dominating turbulence distinctly exhibit a more perpendicular alignment. The comparison between observations and simulations suggests that turbulence could play a more important role than B fields in the formation of condensations in the context of clustered massive star formation, contradicting the prediction of classical magnetically regulated models. Moreover, we find a possibly turbulence-induced preferential misalignment between the B field and rotation axis of condensations, which may potentially reduce the magnetic braking efficiency and facilitate the formation of large protostellar disks. Our findings indicate that turbulence could be critical in determining the initial stellar properties.

astro-ph.GA↗
arXiv2026

CoCA: Step-level Reward for Free in RL-based T2I Diffusion Model Fine-tuning

Recent advances in text-to-image (T2I) diffusion model fine-tuning leverage reinforcement learning (RL) to align generated images with learnable reward functions. The existing approaches reformulate denoising as a Markov decision process for RL-driven optimization. However, they suffer from reward sparsity, receiving only a single delayed reward per generated trajectory. This flaw hinders precise step-level attribution of denoising actions, undermines training efficiency. To address this, we propose a simple yet effective credit assignment framework that dynamically distributes dense rewards across denoising steps. Specifically, we track changes in cosine similarity between intermediate and final images to quantify each step's contribution on progressively reducing the distance to the final image. Our approach avoids additional auxiliary neural networks for step-level preference modeling and instead uses reward shaping to highlight denoising phases that have a greater impact on image quality. Our method achieves 1.25 to 2 times higher sample efficiency and better generalization across four human preference reward functions, without compromising the original optimal policy.

cs.CV↗
arXiv2026

SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning

On-policy distillation (OPD) offers a promising way to transfer reasoning capabilities from stronger teacher models, but applying it to long-context reasoning teachers and short-context students introduces practical challenges, including tokenizer mismatch, teacher-student distribution mismatch, response length explosion, and training instability. In this work, we study this setting by transferring proof-reasoning capabilities from the long-context reasoning model SU-01 to short-context student models. To handle tokenizer differences, we perform OPD in a shared text space and align only tokens that occupy identical text spans under the student and teacher tokenizers. To mitigate the problem of excessive generation length and frequent truncation, we introduce a student reference KL loss and mask the advantages of special termination tokens such as and <|im_end|>. This strategy constrains the student from drifting excessively from its initial policy, thereby mitigating the teacher-student distribution mismatch problem and fostering steady length growth. Experiments on both same-family and different-family student models, including Qwen3, Qwen3.5, Intern-S2, GLM-4.7, Gemma-4, show consistent gains in mathematical reasoning, especially natural-language math proving. Notably, Intern-S2-Preview improves by 21.2 points on ProofBench, reaching 55.2 and surpassing Gemini-2.5-Pro. It also improves on science benchmarks such as HLE and HiPhO, suggesting that OPD transfers reasoning capabilities that generalize beyond the mathematical training domain.

cs.CL↗
arXiv2026

Face-D(^2)CL: Multi-Domain Synergistic Representation with Dual Continual Learning for Facial DeepFake Detection

Facial forgery techniques are advancing rapidly, posing severe threats to public trust and information security while imposing higher demands on the continual adaptation of DeepFake detection models. Although continual learning enables models to adapt to emerging forgery methods, existing approaches still face two key bottlenecks. On the one hand, they lack sufficient feature representation capacity to handle increasingly diverse and complex forgery traces. On the other hand, continual adaptation to new forgery distributions leads to severe catastrophic forgetting of prior knowledge, which substantially degrades detection performance. To address these issues, we propose Face-D(^2)CL, a framework for facial DeepFake detection. It leverages multi-domain synergistic representation to fuse spatial and frequency-domain features, enabling comprehensive capture of diverse forgery traces. Additionally, it employs a dual continual learning mechanism that combines Real/Fake-aware Elastic Weight Consolidation (RF-EWC) and Domain-wise Orthogonal Gradient Constraint (D-OGC). RF-EWC distinguishes the parameter importance for real versus fake samples, while D-OGC ensures that updates to task-specific expert modules do not interfere with previously learned knowledge. This synergy allows the model to achieve a dynamic balance between robust anti-forgetting capabilities and agile adaptability to emerging facial forgery paradigms, all without relying on historical data replay. Extensive experiments demonstrate that our method surpasses current state-of-the-art (SOTA) approaches in both stability and plasticity, achieving a 60.7% relative reduction in the average detection error rate. On unseen forgery domains, it further improves the average detection AUC by 7.9% compared to the current SOTA method.

cs.CV↗
arXiv2026

A Survey of Large Audio Language Models: Generalization, Trustworthiness, and Outlook

Advances in Large Language Models (LLMs) have paved the way for Multimodal Large Language Models (MLLMs). Among these, Large Audio Language Models (LALMs) are essential for realizing universal auditory intelligence. Despite their remarkable performance, the escalation of LALMs' capabilities has significantly outpaced the development of systemic frameworks to ensure their trustworthiness. This survey provides a comprehensive investigation into the endogenous mechanisms of LALMs, detailing the architectural innovations and alignment algorithms that facilitate emergent reasoning. Specifically, we analyze how the transition to unified end-to-end frameworks and the integration of continuous acoustic signals expand the attack surface. To rigorously evaluate the risks within these paradigms, we establish a comprehensive taxonomy of trustworthiness, categorizing critical vulnerabilities such as cross-modal jailbreaking, latent acoustic backdoors, and biometric privacy leakage. We review the state-of-the-art LALMs through six analytical pillars: hallucination, robustness, safety, privacy, fairness, and authentication. The pronounced imbalance between a mature offensive landscape and underdeveloped defenses highlights persistent trustworthiness gaps and multidimensional risks in audio-centric intelligence. Finally, we propose a roadmap advocating for ``Defense-in-Depth'' architectures, causal auditory world modeling, and intrinsic representation engineering to support the development of more reliable and trustworthy audio intelligence. Our project has been uploaded to GitHub https://github.com/Kwwwww74/Awesome-Trustworthy-AudioLLMs.

cs.SD↗
arXiv2026

TI-StegoAlign: Channel-Guided Post-Training for Generative Text Steganography under Tokenization Inconsistency

Generative text steganography enables LLM agents to exchange secret information through task-relevant messages. Yet most methods evaluate recovery on sender-side tokens, whereas the receiver observes only surface text. Detokenization and receiver-side retokenization can alter token boundaries, desynchronize coding states, and cause such evaluation to overestimate receiver-side recovery. Existing remedies rely on inference-time filtering or verification, correcting individual outputs without adapting the generation policy to the receiver-side channel. To address these limitations, we propose TI-StegoAlign, a channel-guided post-training framework. The Bit-Consistent Supervised Objective (BCSO) enlarges local coding margins at realized sender-side embedding positions. Channel-Conditioned Preference Optimization (CCPO) then aligns complete stegotexts using receiver-realistic recovery, text quality, and anti-steganalysis feedback. TI-StegoAlign updates only LoRA parameters and requires no tokenization-specific correction during communication. Experimental results show 100% receiver bit accuracy. Compared with the strongest baselines, TI-StegoAlign achieves a 21.6% reduction in normalized perplexity deviation and a 6.3% relative improvement in anti-steganalysis performance.

cs.CR↗
arXiv2026

Direction-adaptive Mamba: Spatial-Frequency Dual-Domain Collaborative Learning for PolSAR Image Classification

Deep learning dominates polarimetric synthetic aperture radar (PolSAR) image classification, with Mamba architectures serving as favorable backbones due to linear complexity and strong global modeling capacity. However, existing PolSAR Mamba methods have two critical flaws: pure spatial processing discards fine-grained edges and textures, and fixed scanning patterns fail to model direction-variant anisotropic scattering and weak boundaries essential for PolSAR physical analysis. This work proposes DA-Mamba, a direction-adaptive Mamba framework with dual-domain collaborative learning for PolSAR classification. Equipped with an edge-aligned direction-adaptive scanning scheme, DA-Mamba captures long-range spatial dependencies and accurate boundary details. It adopts the Non-Subsampled Contourlet Transform (NSCT) to separate PolSAR data into low-frequency global components and multi-directional high-frequency subbands, extracting anisotropic structural features from high-frequency information while preserving global context via low-frequency branches. A dual-domain collaborative learning module further integrates spatial scattering and frequency-domain representations to strengthen feature discriminability. Evaluated on three real-world PolSAR datasets, DA-Mamba surpasses state-of-the-art methods, verifying the efficacy of the proposed adaptive scanning and dual-domain fusion designs. Code will be publicly available.

eess.IV↗
arXiv2026

WGDnet: Wishart-guided Geometric-aware Deep Network for PolSAR Image Classification

Polarimetric Synthetic Aperture Radar (PolSAR) classification underpins all-weather Earth observation. Conventional Wishart methods depend on rigid handcrafted operators with limited adaptability, while mainstream deep networks ignore PolSAR native Wishart scattering statistics. Additionally, fixed convolution windows fail to capture multi-scale, multi-directional terrain patterns, harming boundary detection and small-object characterization. To mitigate these drawbacks, we propose WGDNet, a Wishart-guided geometric-aware deep network. It integrates three core designs: (1) learnable Wishart convolutions with directional kernels for multi-scale statistical edge feature extraction; (2) an orientation-prior aggregation module that estimates dominant local directions and confidences to refine directional Wishart outputs adaptively; (3) GAnet, a scale-direction adaptive geometric-aware convolution that dynamically reshapes sampling grids to model anisotropic terrain and retain fine details. Our contributions lie in learnable Wishart statistical modeling, orientation-prior feature aggregation, and geometry-adaptive convolution. Evaluations across four real PolSAR datasets verify WGDNet surpasses existing state-of-the-art approaches in classification accuracy and boundary fidelity.

cs.CV↗