arXiv subjects

Zhiqiang Shen

Publications and source records attributed to Zhiqiang Shen.

At least 19 recordsLinked to original sources

arXiv2026

ResQAdam: Residual-Preserving Adaptive Optimization for Memory-Efficient Quantization Aware Training

Quantization-aware training (QAT) is the most reliable way to obtain accurate low-bit large language models (LLMs), but training with AdamW keeps two dense FP32 states for every weight matrix. We find that low-rank memory-efficient optimizers such as GaLore and LDAdam degrade substantially in QAT. A considerable part of the QAT momentum lies outside its top-r singular subspace, and projection-based methods discard this part at every step. We further observe that this residual remains useful when only its sign is kept. Based on these observations, we propose ResQAdam. It keeps a single BF16 momentum, applies Adam only to the r x r coordinates of the principal subspace, and adds a normalized sign update computed from the residual. We evaluate ResQAdam with two from-scratch QAT methods and three QAT fine-tuning pipelines, with weights quantized down to 2 bits. It reduces optimizer-state memory by 71% to 74% relative to AdamW, stays close to AdamW in most settings, and lowers training throughput by less than 2%.

cs.LG↗

arXiv2026

DELTA: Robustly Training Label-Conditional Diffusion Models with Weak Annotations

While label-conditional diffusion models have recently exhibited remarkable generative capabilities, their success heavily relies on massive, cleanly labeled datasets. In practice, categorical supervision is rarely perfect: it is often corrupted by noise, clouded by ambiguity, or partially missing. Training directly on such weak annotations severely degrades generation quality, and existing robust methods offer only fragmented solutions that depend on scarce auxiliary priors. To overcome these limitations, we propose DELTA, a unified framework that robustly trains Diffusion modELs across all three weak annoTAtion types. By treating the unknown true label as a latent variable, DELTA optimizes a principled variational objective that jointly recovers the clean data distribution and infers the true label posteriors. To make this joint training computationally tractable, we further introduce a median-centered timestep sampling strategy that efficiently concentrates estimation where the diffusion process is most informative. Extensive experiments on synthetic and real-world benchmarks demonstrate that DELTA produces high-fidelity, class-consistent samples across diverse weak supervision scenarios, outperforming specialized baselines while demanding strictly less prior knowledge.

cs.LG↗

arXiv2026

Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models

Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable 10.12$\times$ and 5.12$\times$ speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance. Our code is available at https://github.com/VILA-Lab/dSpark.

cs.CL↗

arXiv2026

Extreme Compression of Per-Layer Embeddings via Cross-Layer Latent Vector Quantization

Scaling embedding capacity alongside Transformer computation is an emerging trend in language models. This trend makes high compression rates increasingly useful. Hence, we study extreme compression of Per-Layer Embeddings (PLE), which follow this trend and account for 35.2% of the parameters of Gemma 4 E4B. At such rates, conventional quantizers give each token-layer vector its own code of only a few bytes. We instead develop OPTIC (Output-Path-fitted, Token-shared Indexed Codes), which stores one quantized code tuple per token and decodes it at every layer with a layer-specific affine decoder. Codes and decoders are fitted to the PLE output path rather than to the stored vectors. At about $100\times$ PLE compression of bfloat16 (BF16) E4B, OPTIC retains an aggregate 97.8% of the uncompressed model's performance across seven benchmarks, whereas product quantization, the strongest of six conventional baselines, retains 66.3%. At $594\times$ and $1,002\times$ PLE compression of BF16 E4B, OPTIC retains 90.5% and 84.1%. On the official E4B release with 4-bit linear weights and 16-bit activations (W4A16), OPTIC retains 96.9% at about $100\times$ PLE compression. During inference at about $100\times$ PLE compression, BF16 E4B uses 29.7% less GPU memory than with uncompressed PLE at the same key-value cache size, while running at most 0.40% slower.

cs.LG↗

arXiv2026

Pushing the Frontier of Black-Box LVLM Attacks via Fine-Grained Detail Targeting

Black-box adversarial attacks on Large Vision-Language Models (LVLMs) are challenging due to missing gradients and complex multimodal boundaries. Prior state-of-the-art transfer-based approaches such as M-Attack match local crops between source and target images, but this induces high-variance, nearly orthogonal gradients across iterations. We attribute this instability to (i) ViT translation sensitivity, which yields spike-like gradients, and (ii) a structural asymmetry between source and target crops. We reformulate local matching as an asymmetric expectation over source transformations and target semantics, and introduce M-Attack-V2 based on this formulation. Multi-Crop Alignment (MCA) averages gradients from multiple local views per iteration, while Auxiliary Target Alignment (ATA) replaces aggressive target augmentation with a small set of semantically correlated auxiliary images. Patch Momentum replays historical crop gradients, and a patch-diverse surrogate ensemble (PE+) strengthens transferable directions. M-Attack-V2 raises attack success rates over M-Attack from 40% to 61% on GPT-6.1 Sol, from 40% to 60% on GPT-6 Astra, and from 85% to 97% on Claude Sonnet 5.5. Code and data are publicly available at: https://github.com/VILA-Lab/M-Attack-V2.

cs.LG↗

arXiv2026

LLMSurgeon: Diagnosing Data Mixture of Large Language Models

The pretraining data mixture of Large Language Models (LLMs) constitutes their "digital DNA", shaping model behaviors, capabilities, and failure modes. Yet this composition is rarely disclosed, making post-hoc auditing of data combination or provenance difficult. In this work, we formalize Data Mixture Surgery (DMS): given only generated text from a target LLM, estimate the domain-level distribution of its pretraining corpus under a predefined taxonomy. We propose $\textbf{LLMSurgeon}$, a strong framework that casts DMS as an inverse problem under the label-shift assumption. Rather than directly aggregating classifier outputs, LLMSurgeon estimates a calibrated $\textit{soft}$ confusion matrix and solves a constrained inverse problem to correct systematic domain confusion and recover the latent mixture prior. To evaluate, we introduce $\textbf{LLMScan}$, a recipe-verifiable evaluation suite built from open-source LLMs with transparent pretraining mixtures. Across LLMScan, LLMSurgeon recovers domain mixtures with high fidelity under fixed protocols. Our work presents a practical, post-hoc approach for auditing the digital DNA of foundation models without access to their training data.

cs.CL↗

arXiv2026

Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models

Large Language Models (LLMs) with billions of parameters are prime targets for network pruning, which removes some model weights without hurting performance. Prior approaches such as magnitude pruning, SparseGPT, and Wanda either concentrated solely on weights or integrated weights with activations for sparsity. However, they overlooked the informative gradients derived from pretrained LLMs. In this paper, we present a novel sparsity-centric pruning method for pretrained LLMs, termed Gradient-based Language Model Pruner (GBLM-Pruner). GBLM-Pruner leverages the first-order term of the Taylor expansion, operating in a training-free manner by harnessing properly normalized gradients from a few calibration samples to determine the pruning metric, and substantially outperforms competitive counterparts like SparseGPT and Wanda in multiple benchmarks. Intriguingly, by incorporating gradients, unstructured pruning with our method tends to reveal some structural patterns, which mirrors the geometric interdependence inherent in the LLMs' parameter structure. Additionally, GBLM-Pruner functions without any subsequent retraining or weight updates to maintain its simplicity as other counterparts. Extensive evaluations on LLaMA-series, Qwen-Series and Deepseek models across zero-shot, STEM-Reasoning and Coding benchmarks show that GBLM-Pruner surpasses magnitude pruning and Wanda by significant margins and performs competitively against SparseGPT which involves intensive weight updates. We further extend our approach on Vision Transformer and Vision Language Models. Our code and models are available at https://github.com/VILA-Lab/GBLM-Pruner.

cs.CL↗

arXiv2026

WUSPruner: Second-Moment-Optimal Transforms for Structured Pruning

Structured pruning is typically formulated as identifying which model dimensions to remove, often within the model's original representation. However, the effect of coordinate deletion fundamentally depends on the basis in which pruning is performed. This raises a more fundamental question: how should the basis be chosen specifically for coordinate deletion? We introduce WUSPruner, a transform-based framework for structured pruning of large language models. We formulate basis selection as a local linear reconstruction problem that directly minimizes the reconstruction loss induced by pruning. Inspired by the second-moment transformation in WUSH, we consider a WUS transform constructed jointly from activation and weight statistics. We show that under positive-definite second moments, WUS transform minimizes the local reconstruction loss over all invertible changes of basis. In the resulting basis, the activation and weight second moments are simultaneously diagonalized and balanced, causing the deletion error to decompose into independent spectral contributions and yielding a principled criterion for coordinate selection. Guided by this result, we develop a scalable pruning method based on blockwise WUS transformations and compact input projections. Experiments on Llama and Qwen models from 8B to 72B parameters show consistent improvements over structured pruning baselines across language modeling, knowledge, commonsense reasoning, mathematics, and code generation benchmarks, while also delivering practical reductions in inference latency and memory usage.

cs.LG↗

arXiv2026

GA-SVD: Generation-aware Singular Value Decomposition for LLM Compression

Large language models (LLMs) demonstrate strong generalization across a broad range of reasoning and generation tasks, but their large parameter counts impose substantial memory and computation costs, hindering deployment in practical resource-constrained settings. One-shot low-rank compression is a promising solution since it reduces model size and inference cost effectively. However, existing SVD-based methods primarily target multiple-choice benchmarks and often fail to preserve performance on generation tasks. Moreover, their rank allocation strategies are typically heuristic. We propose GA-SVD, a generation-aware low-rank compression framework for LLMs. First, we introduce Generation-Aware Weight Whitening (GAWW), which leverages gradients from on-policy knowledge distillation to construct whitening transformations aligned with generative behavior. Second, we develop an evolutionary global rank allocation method that searches for effective layer-wise rank allocation under a given global compression ratio. Across extensive experiments on Mistral-7B, Qwen3 series, and Llama-3.3 models, evaluated on both multiple-choice and generative benchmarks, GA-SVD consistently outperforms existing SVD-based baselines and achieves state-of-the-art (SoTA) results. We further demonstrate scalability by applying our method to models with up to 70B parameters.

cs.LG↗

arXiv2026

Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems

Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its architecture by analyzing the publicly available source code and comparing it with two independent open-source AI agent systems, OpenClaw and Hermes Agent, that answer many of similar or even the same design questions. Our analysis identifies five human values, philosophies, and needs that motivate the architecture: human decision authority, safety, security, and privacy, reliable execution, capability amplification, and contextual adaptability. We then trace them through thirteen design principles to implementation choices. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. Most of the code, however, lives in the systems around this loop: a permission system with seven modes and an ML-based classifier, a five-layer compaction pipeline for context management, four extensibility mechanisms (MCP, plugins, skills, and hooks), a subagent delegation and orchestration mechanism, and append-oriented session storage. Comparisons with OpenClaw and Hermes Agent show that the same design questions produce different answers across three deployment contexts. Claude Code emphasizes per-action safety, OpenClaw emphasizes perimeter-level access, and Hermes renders per-action approvals across many surfaces. At the runtime layer, Claude Code uses a single CLI loop, OpenClaw embeds the runtime within a gateway control plane, and Hermes uses one process whose role is set by its entry point. At the context and extension layer, Claude Code extends the context window, OpenClaw registers gateway-wide capabilities, and Hermes provides pluggable memory and model backends. We finally identify six open design directions for future agent systems, grounded in recent empirical, architectural, and policy literature.

cs.SE↗

arXiv2026

ActiveSAM: Fast and Accurate Open-Vocabulary Semantic Segmentation with Frozen SAM 3

Segment Anything Model 3 (SAM 3) provides a strong frozen backbone for concept-prompted segmentation, but applying it directly to open-vocabulary semantic segmentation (OVSS) is inefficient: full-resolution decoding is typically run over the entire dataset vocabulary, whereas each image contains only a small active subset of classes. We introduce ActiveSAM, a training-free inference framework that turns SAM 3 into an active-vocabulary segmenter. ActiveSAM first canonicalizes and expands class prompts, then uses evidence-proportional grounding to estimate an image-conditioned active set from a low-resolution presence preview. Only retained prompts receive full-resolution mask prediction, using bucketed prompt multiplexing with the frozen SAM 3 decoder. The preview stage uses only class-presence evidence and skips unnecessary segmentation-head computation. To resolve overlapping concept responses, exclusive concept decoding compares each pixel's joint score vector with class signatures estimated once per vocabulary from unlabeled images. ActiveSAM requires no weight updates, no oracle class-presence labels and no per-dataset hyperparameter tuning. Across eight OVSS benchmarks, ActiveSAM improves the speed-accuracy tradeoff of training-free open-vocabulary semantic segmentation, outperforming the current state-of-the-art SegEarth-OV3 by +2.1 mIoU on average while running much faster, with 7.3-12.2x speedups on large-vocabulary datasets. ActiveSAM also achieves the highest accuracy under image corruptions that simulate real-world distribution shift, making it well-suited for deployment in noisy-input domains such as autonomous driving and embodied AI. Code is available at https://github.com/VILA-Lab/ActiveSAM

cs.CV↗

arXiv2026

OD3: Optimization-free Dataset Distillation for Object Detection

Training large neural networks on large-scale datasets requires substantial computational resources, particularly for dense prediction tasks such as object detection. Although dataset distillation (DD) has been proposed to alleviate these demands by synthesizing compact datasets from larger ones, most existing work focuses solely on image classification, leaving the more complex detection setting largely unexplored. In this paper, we introduce OD3, a novel optimization-free data distillation framework specifically designed for object detection. Our approach involves two stages: first, a candidate selection process in which object instances are iteratively placed in synthesized images based on their suitable locations, and second, a candidate screening process using a pre-trained observer model to remove low-confidence objects. We perform our data synthesis framework on MS COCO and PASCAL VOC, two popular detection datasets, with compression ratios ranging from 0.25% to 5%. Compared to the prior solely existing dataset distillation method on detection and conventional core set selection methods, OD3 delivers superior accuracy, establishes new state-of-the-art results, surpassing prior best method by more than 14% on COCO mAP50 at a compression ratio of 1.0%. Code is available at: https://github.com/VILA-Lab/OD3.

cs.CV↗

arXiv2026

Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value (KV) caching and scalable parallel decoding mechanisms. Existing acceleration methods typically study KV caching and parallel decoding in isolation, overlooking the I/O bottlenecks that arise when cache reuse and parallel token verification are jointly applied. In this work, we introduce $\textbf{Flash-dLLM}$, a training-free inference acceleration framework for fast and memory-efficient dLLMs. Flash-dLLM first identifies GPU memory I/O as a dominant bottleneck in KV-cache-enabled dLLM inference and addresses it with an I/O-aware fused KV-cache kernel that reduces redundant memory movement. Building on this optimized cache mechanism, Flash-dLLM further proposes an efficient KV-cache-driven draft-and-verify decoding strategy, where the dLLM itself serves as both drafter and verifier without requiring an auxiliary model. This unified design enables faster decoding while preserving generation quality and improving scalability to longer sequences and larger batch size. Extensive experiments on mathematical reasoning and code-generation benchmarks demonstrate that Flash-dLLM consistently outperforms existing state-of-the-art dLLM acceleration methods in both inference speed and memory efficiency. In particular, it achieves $5.1\times$ and $11.0\times$ speedups over prior strongest baseline Elastic-Cache on GSM8K and HumanEval, respectively.

cs.CL↗

arXiv2026

ALMA Polarization Study of the Magnetic Fields in Two Massive Clumps in the 20 km s$^{-1}$ Cloud of the Central Molecular Zone

We present the Atacama Large Millimeter/submillimeter Array (ALMA) observations of linearly polarized 870 $μ$m continuum emission at a resolution of $\sim$0.2$^{\prime\prime}$ (2000 au) toward the two massive clumps, Clump 1 and Clump 4, in the 20 km s$^{-1}$ cloud. The derived magnetic field strengths for both clumps range from $\sim$0.3 to 3.1 mG using the Angular Dispersion Function (ADF) method. The magnetic field orientations across multiple scales suggests that the magnetic field dominates at the cloud scale, whereas gravity likely governs the structures at the core (0.01$-$0.1 pc) and condensation ($\le$ 0.01 pc) scales. Furthermore, the study on the angular difference between the orientations of the local gravity gradient and the magnetic field suggests that the gravity predominantly governs the dynamics in the diffuse regions, while both gravity and star formation feedback become increasingly significant within the dense regions. The ratio of the magnetic field tension force $F_\mathrm{B}$ to the gravitational force $F_\mathrm{G}$ suggests that the magnetic field may provide some support against gravity, but it is insufficient to prevent gas from infalling toward the dense cores.

astro-ph.GA↗

arXiv2026

FigMirror: Ground It, Code It, Plot It

Converting scientific figures into executable code has gained increasing attention, yet existing methods primarily focus on reproducing the reference figure itself. A more practical setting is to plot new data while preserving the visual style of a reference figure (e.g., color scheme and typography). Prior approaches mimic the reference through pixel-level optimization and struggle to carry its style to new data. We show that the key to this task lies in the coordinate grounding and coding capabilities present in modern computer-use models. We propose FigMirror, an agentic framework that unlocks these capabilities through Grounded Measurement, which locates visual elements by coordinates and measures their properties through executable code. We further introduce PlotTwin-Bench, an expert-curated benchmark with fine-grained code and image-level style metrics. Experiments show that FigMirror consistently outperforms existing methods on reference-conditioned style transfer. All plots in this paper are generated by FigMirror, except those produced by other methods for comparison. Our code and data are available at: https://github.com/VILA-Lab/FigMirror.

cs.CV↗

arXiv2026

OH Line Detections in Southern Galaxies of the IRAS Revised Bright Galaxy Sample

We present a systematic study of OH main-line emission and absorption in 186 southern galaxies from the IRAS Revised Bright Galaxy Sample, using archival MeerKAT snapshot data. OH features are detected in 38 galaxies, including eight with OH maser emission (three new) and 30 showing OH absorption, mostly unreported previously. Four absorption systems exhibit weak OH emission superposed on strong absorption. OH-emitting regions are generally more compact than the associated radio continuum. Most absorption profiles are well fit by two Gaussian components (1667 and 1665 MHz), with an average integrated line ratio of $\sim$1.5. LIRGs show an OH emission detection rate of ~13\%, versus significantly lower rates in non-LIRGs. For sources with radio continuum flux densities >20 mJy, OH absorption detection rates reach ~36\% (LIRGs) and ~27\% (non-LIRGs), while no OH absorption features were detected among sources with lower radio continuum flux densities. This suggests that sufficient background continuum is likely an important factor for the detection of OH absorption. Detected OH emitters follow the empirical $L_{\rm OH}$--$L_{\rm FIR}$ relation, consistent with far-infrared pumping, while non-detections show upper limits below the relation. No significant differences are found between OH absorbers and non-detections in infrared luminosity or radio continuum compactness. Stacked spectra of non-detections reveal no significant OH features, suggesting that sensitivity and orientation alone do not fully explain the absence of absorption. In contrast, mid-infrared colors (e.g., W2--W3) and q_TIR differ between the two populations. OH absorption galaxies occupy an intermediate regime in L_HCN/L_CO between OH megamasers and non-detections, implying that OH absorption detectability is linked to dense molecular gas conditions, with extreme star formation potentially suppressing its occurrence.

astro-ph.GA↗

arXiv2026

Beyond Sgr A* and M87*: Sub-Microarcsecond Black Hole Shadow Detection via Lunar-based Extremely Long Baseline Interferometry

The 1.3 mm ground-based very long baseline interferometry (VLBI) array Event Horizon Telescope (EHT), is limited by Earth's diameter, restricting horizon-scale imaging to only M87* and Sgr A*. Extending baselines to the Moon would achieve ~0.7 microarcsecond angular resolution at 230 GHz, enabling shadow detection for many more supermassive black holes (SMBHs). The concept is motivated by space VLBI missions and lunar exploration, including the ongoing Lunar Orbit VLBI EXperiment (LOVEX) aboard QueQiao-2 (Chang'E-7) and the planned International Lunar Research Station (ILRS). We assess shadow detectability for 31 SMBHs with predicted large angular sizes, assuming optically thin emission at 230 GHz, exploring different telescope locations and antenna sizes. Assuming a telescope at the lunar antipode, we simulate the Moon-Earth (u,v) coverage and show that sources near the Moon's orbital plane yield projected baselines spanning a wide range, enabling sampling of the first visibility null - a key shadow signature. Using a geometric ring model, we identify six shadow-detectable candidates: M104, NGC 5077, and NGC 1052 are detectable with a 5 m lunar-based telescope; PGC 049940 with 10 m; NGC 524 with 20 m; and NGC 5252 with 40 m. If additional space telescopes fill the baseline coverage gaps between Moon and Earth, 14 candidates are detectable for the n=1 photon-ring region with a lunar-based telescope up to 40 m. These results provide a clear scientific and technical motivation for lunar-based telescopes in future black hole shadow studies.

astro-ph.GA↗

arXiv2026

DARTree: Speculative Diffusion Decoding with Autoregressive Draft Trees

Speculative decoding losslessly accelerates autoregressive language models by verifying multiple draft tokens in parallel. Diffusion-based drafters further reduce proposal latency by predicting an entire token block in parallel, but their position-wise distributions are marginal rather than conditioned on tokens selected along each draft path. Existing recurrent correction incorporates causal information along a single draft chain, whereas diffusion-based tree construction broadens candidate coverage without carrying this correction along individual branches. We introduce DARTree, a training-free speculative decoding method that extends a pretrained AR correction head from chains to trees. DARTree first constructs a fixed-width candidate tree by expanding and scoring all nodes at each depth in a single batch, and then only applies best-first pruning to select the verification tree, decoupling AR-head inference from sequential heap operations. Across seven math, code, and chat benchmarks, DARTree achieves the highest average acceptance length and speedup in all four model--temperature configurations, accepting up to 12.97 tokens per verification round, 98.6\% more than DFlash and 27.9\% more than Domino in the same setting, and reaching up to 9.73$\times$ lossless speedup over locally measured autoregressive decoding.

cs.LG↗