arXiv subjects

Tao Zhang

Publications and source records attributed to Tao Zhang.

At least 127 records · Page 7Linked to original sources

arXiv2026

Determinism of Randomness: Prompt-Residual Seed Shaping for Diffusion Generation

Diffusion models start generation from an isotropic Gaussian latent, yet changing only the random seed can lead to large differences in prompt faithfulness, composition, and visual quality. We study this seed sensitivity through the semantic map from initial noise to generated meaning. Although the sampling flow is locally invertible, the subsequent semantic projection is many-to-one, inducing a degenerate pullback semi-metric on the latent space: most local directions are nearly semantic-invariant, while semantic-sensitive variation is concentrated in a much smaller horizontal subspace. This provides an explanatory geometric view of the seed lottery. Motivated by this view, we introduce a training-free prompt-residual seed-shaping procedure. Rather than claiming to recover the exact horizontal space, the method uses a single high-noise cold-start prompt residual as a model-coupled proxy, injects only its tangential component, and retracts the seed to the original Gaussian radius shell. This keeps the initialization prior-compatible while adding only one conditional/unconditional probe before standard sampling. Across multiple generation benchmarks, the method improves alignment and quality metrics over standard sampling, supporting both the practical value of the proxy and the explanatory relevance of semantic anisotropy.

cs.CV↗

arXiv2026

A proof of purely singular splitting conjecture

A set $M$ of nonzero integers is said to split a finite abelian group $G$ if there exists a subset $S\subseteq G$ such that $M\cdot S = G\setminus\{0\}$. Such a splitting is called purely singular if every prime divisor of $|G|$ divides some element of $M$. In 1995, Woldar \cite{W1995} conjectured that the finite abelian groups admitting a purely singular splitting by the set $\{1,2,\dots,k\}$ are precisely the cyclic groups of orders $1$, $k+1$, and $2k+1$. In this paper, we prove this conjecture.

math.CO↗

arXiv2026

Utility-Aware Progressive Inference over UDP Packet Blocks for Emergency Communications

Emergency communications increasingly rely on remote visual inference for timely hazard detection under stringent bandwidth and latency constraints. However, conventional UDP-based visual delivery typically performs inference only after the full payload has been received, even though partially received packet blocks may already contain sufficient task-relevant evidence for reliable decision making. This paper proposes a utility-aware progressive inference framework for emergency communications, which operates directly on UDP packet blocks and determines when sufficient task value has been accumulated for early hazard recognition. Specifically, the sender estimates packet-level decision utility as lightweight control metadata, while the receiver progressively updates partial observations, accumulates the utility of received packets, and triggers an early stop once the normalized utility exceeds a prescribed threshold. Experiments on a fire-scene detection dataset show that, at the main operating point, the proposed method reduces the average packet budget by 34.2% and the decision delay by 1209.17 ms while retaining 91.5% of the full-reception match rate. The method also maintains its advantage over the stability-based baseline under moderate packet loss and different packet-arrival orders. These results demonstrate that packet-level utility provides an effective basis for communication-efficient and delay-aware hazard recognition over UDP-based emergency links.

eess.SP↗

arXiv2026

In-Network Artificial Computing Enhanced Light Model-Switching for Emergency Communications Networks

Emergency communications networks require in-network intelligence for timely traffic handling under dynamic demands and runtime constraints. In these environments, packets may need different inference behaviors, and conventional model replacement via control-plane updates is too slow for responsive operation. We propose an in-network artificial computing framework with lightweight model-switching, where multiple Binary Neural Network (BNN) models are kept resident within a shared execution framework. Packet metadata selects the active model at packet granularity with O(1) selection cost. A fixed 1024-byte payload is aligned with x86 AVX-512, enabling efficient memory access. The framework is realized on an eBPF/XDP + AF_XDP stack. Experimental results show that the system sustains 1.894 Mpps with a 0.528 us inference latency, while model selection adds only 0.005 us. Our results demonstrate that different resident models induce distinct packet-processing behaviors, that scaling to 16 slots preserves low switching overhead, and that online model switching completes without wrong-verdict packets. These results show the practicality of lightweight in-network artificial computing on commodity hardware.

cs.NI↗

arXiv2026

Using large language models for embodied planning introduces systematic safety risks

Large language models are increasingly used as planners for robotic systems, yet how safely they plan remains an open question. To evaluate safe planning systematically, we introduce DESPITE, a benchmark of 12,279 tasks spanning physical and normative dangers with fully deterministic validation. Across 23 models, even near-perfect planning ability does not ensure safety: the best-planning model fails to produce a valid plan on only 0.4% of tasks but produces dangerous plans on 28.3%. Among 18 open-source models from 3B to 671B parameters, planning ability improves substantially with scale (0.4-99.3%) while safety awareness remains relatively flat (38-57%). We identify a multiplicative relationship between these two capacities, showing that larger models complete more tasks safely primarily through improved planning, not through better danger avoidance. Three proprietary reasoning models reach notably higher safety awareness (71-81%), while non-reasoning proprietary models and open-source reasoning models remain below 57%. As planning ability approaches saturation for frontier models, improving safety awareness becomes a central challenge for deploying language-model planners in robotic systems.

cs.AI↗

arXiv2026

Training cell stress patterns in 3D cellular packings

The task of learning patterns is typically associated with systems that update parameters on fixed architectures, such as neural networks, where learning proceeds through continuous optimization. Here, we demonstrate that pattern learning can also emerge in reconfigurable cellular tissue, where both mechanical parameters and network topology evolve. Using a three-dimensional vertex model, we show that cellular packings can be trained to realize prescribed cell stress patterns through a contrastive learning algorithm to update hidden-cell shape indices. We find that learning is intrinsically collective, requiring coordinated, system-wide parameter adjustments, with learnability governed by an interplay between mechanical state, capacity, and training protocol. In particular, the rigidity of the tissue controls an effective exploration-exploitation tradeoff: fluid-like regimes enhance exploration through cellular rearrangements, while rigid regimes constrain dynamics and favor exploitation of existing configurations. These rearrangements introduce discontinuous learning dynamics, enabling the system to transition between distinct local minima in the cost function landscape. As the ratio of target cells to the total number cells in the packing or constraint load increases, learning becomes slower, more heterogeneous, and increasingly dependent on rare rearrangements that allow escape from geometrically constrained states. Finally, training cells in sequence, in contrast to parallel protocols, provides an alternative route that can be more robust but generally takes longer to train for the constraint loads studied. These results suggest a learning phase diagram governed by constraint load, cell packing rigidity, and training protocol. By enabling the training of localized internal states, this work positions tissues not only as adaptive materials, but as nonconventional AI platforms.

cond-mat.dis-nn↗

arXiv2026

Report of the 5th PVUW Challenge: Towards More Diverse Modalities in Pixel-Level Understanding

This report summarizes the objectives, datasets, and top-performing methodologies of the 2026 Pixel-level Video Understanding in the Wild (PVUW) Challenge, hosted at CVPR 2026, which evaluates state-of-the-art models under highly unconstrained conditions. To provide a comprehensive assessment, the 2026 edition features three specialized tracks: the MOSE track for tracking objects within densely cluttered and severely occluded scenarios; the MeViS-Text track for localizing targets via motion-focused linguistic expressions; and the newly inaugurated MeViS-Audio track, which pioneers acoustic-driven object segmentation. By introducing previously unreleased challenging data and analyzing the cutting-edge, multimodal solutions submitted by participants, this report highlights the community's latest technical advancements and charts promising future directions for robust video scene comprehension.

cs.CV↗

arXiv2026

RailVQA: A Benchmark and Framework for Efficient Interpretable Visual Cognition in Automatic Train Operation

As Automatic Train Operation (ATO) advances toward GoA4 and beyond, it increasingly depends on efficient, reliable cab-view visual perception and decision-oriented inference to ensure safe operation in complex and dynamic railway environments. However, existing approaches focus primarily on basic perception and often generalize poorly to rare yet safety-critical corner cases. They also lack the high-level reasoning and planning capabilities required for operational decision-making. Although recent Large Multi-modal Models (LMMs) show strong generalization and cognitive capabilities, their use in safety-critical ATO is hindered by high computational cost and hallucination risk. Meanwhile, reliable domain-specific benchmarks for systematically evaluating cognitive capabilities are still lacking. To address these gaps, we introduce RailVQA-bench, the first VQA benchmark for cab-view visual cognition in ATO, comprising 20,000 single-frame and 1,168 video based QA pairs to evaluate cognitive generalization and interpretability in both static and dynamic scenarios. Furthermore, we propose RailVQA-CoM, a collaborative large-small model framework that combines small-model efficiency with large-model cognition via a transparent three-module architecture and adaptive temporal sampling, improving perceptual generalization and enabling more efficient reasoning and planning. Experiments demonstrate that the proposed approach substantially improves performance, enhances interpretability, improves efficiency, and strengthens cross-domain generalization in autonomous driving systems. Code and datasets will be available at https://cybereye-bjtu.github.io/RailVQA.html.

cs.CV↗

arXiv2026

How Do Answer Tokens Read Reasoning Traces? Self-Reading Patterns in Thinking LLMs for Quantitative Reasoning

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

cs.CL↗

arXiv2026

Mapping-based Hard-constrained Physics-Informed Neural Networks for unbounded wave problems

The aim of this paper is to introduce a Mapping-based Hard-constrained Physics-Informed Neural Network (MH-PINN) for efficiently and accurately solving unbounded wave problems. First, we propose a coordinate mapping technique that compactifies the infinite physical domain into a finite computational space. This effectively resolves the sampling difficulties inherent to standard PINNs in unbounded regions. Additionally, it avoids the artificial truncation errors introduced by traditional methods such as perfectly matched layers. Second, we design a physics-based hard-constrained network structure that automatically satisfies both the inner boundary conditions and the far-field radiation conditions. This structure eliminates boundary loss terms, yielding high computational efficiency and fast convergence, which effectively addresses the challenges of high-frequency problems. Third, we introduce an inverse factor correction for boundary coefficients to address the influence of asymptotic factors,which makes the method highly geometrically adaptable. Finally, we present numerical examples covering various acoustic radiation and scattering scenarios as well as elastic dynamics scenarios to demonstrate the efficiency and accuracy of our algorithm.It highlights its potential for broader applications in the field of computational wave dynamics.

math.NA↗

arXiv2026

Graph-based Hierarchical Deep Reinforcement Learning for Deliverable Block Propagation with Optimal Hybrid Cost in Web 3.0

Web 3.0 is envisioned as a decentralized paradigm, where blockchain serves as a core technology for transparent and tamper-proof data management. Among various blockchain architectures, consortium blockchains have emerged as the preferred platform for enterprise-grade Web 3.0. For consortium blockchains, newly generated blocks are generally propagated to all consensus nodes for validation through the gossip protocol. However, gossip-based propagation may introduce substantial message redundancy and tail latency. Moreover, the consensus nodes exhibit heterogeneous availability patterns, and existing block propagation schemes often overlook such temporal constraints. Therefore, the joint optimization of propagation timeliness and delivery coverage remains an open problem. In this paper, we propose a deliverable block propagation optimization framework for consortium blockchain-enabled Web 3.0. We first propose a delivery-aware timeliness metric called Age of Validated Block (AoVB), which excludes block receptions occurring outside the availability window of each consensus node, thereby measuring only actionable synchronization latency. This metric is unified with the block arrival rate into a hybrid cost objective that balances timeliness against delivery. To solve this complex optimization problem, we propose a Graph-based Hierarchical Deep Reinforcement Learning (GHDRL) method, which comprises a graph isomorphism network-based assignment module and a graph attention network-based propagation module. The two modules are optimized jointly under a two-stage training strategy. Numerical results show that GHDRL consistently outperforms all compared schemes across network scales from 50 to 500 peers, achieving up to 19.2% lower hybrid cost than the best-performing neural baseline. Moreover, the model generalizes from 100-peer training instances to 500-peer deployments without retraining.

cs.NI↗

arXiv2026

Evolution of topological phases in atomically thin WTe2 films

Topological materials ranging from topological insulators to semimetals host many novel quantum phenomena including quantum spin Hall effect and topological Fermi arcs. Transitions between these topological phases have attracted much research interest. We performed angle-resolved photoemission spectroscopy (ARPES) on WTe2 ranging from a monolayer to the bulk and reveal the evolution of the electronic structure and the band gap. Notably, the gap observed in the monolayer system is suppressed in the three layers, where the film becomes metallic. Variations in the topological properties with thickness are demonstrated by the first-principles calculations. Topological Z2 invariant is shown to oscillate between 1 and 0 with the addition of layers, originating from the interlayer coupling-induced change in band crossing. The system evolves into a Weyl semimetal when the conduction and valence bands touch near the Fermi level and the topological nature is described by the Chern number. Our findings demonstrate the non-monotonic dependence of topological states on dimensionality and how layer-driven electronic band reconfiguration leads to phase transitions in solids.

cond-mat.mes-hall↗

arXiv2026

Native Hybrid Attention for Efficient Sequence Modeling

Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra & inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single softmax attention operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.

cs.CL↗

arXiv2026

Physically-Guided Optical Inversion Enable Non-Contact Side-Channel Attack on Isolated Screens

Noncontact exfiltration of electronic screen content poses a security challenge, with side-channel incursions as the principal vector. We introduce an optical projection side-channel paradigm that confronts two core instabilities: (i) the near-singular Jacobian spectrum of projection mapping breaches Hadamard stability, rendering inversion hypersensitive to perturbations; (ii) irreversible compression in light transport obliterates global semantic cues, magnifying reconstruction ambiguity. Exploiting passive speckle patterns formed by diffuse reflection, our Irradiance Robust Radiometric Inversion Network (IR4Net) fuses a Physically Regularized Irradiance Approximation (PRIrr-Approximation), which embeds the radiative transfer equation in a learnable optimizer, with a contour-to-detail cross-scale reconstruction mechanism that arrests noise propagation. Moreover, an Irreversibility Constrained Semantic Reprojection (ICSR) module reinstates lost global structure through context-driven semantic mapping. Evaluated across four scene categories, IR4Net achieves fidelity beyond competing neural approaches while retaining resilience to illumination perturbations.

cs.CV↗

arXiv2026

An Argument-Principle Based Stability Assessment Method for Grey-Box DFIG Systems

Considerable efforts have been made to analyze the small-signal stability of doubly fed induction generator (DFIG) systems. However, commercial confidentiality and frequency coupling make the DFIG system a grey-box multiple-input-multiple-output (MIMO) system with highly challenging stability analysis. This paper proposes an Argument-principle based stability assessment method to analyze the stability of the grey-box DFIG system. The frequency sweeping technique is first used to acquire the MIMO model of the black-box device, as well as the determinant of the system's return difference matrix. Then a stability criterion based on the determinant trajectory is presented. This criterion applies to the stability analysis of grey-box MIMO systems without detailed system models. Further, acritical-pole estimation method with trajectory information is developed to assess the dominant mode of the target system. The simulation and hardware-in-loop experiment results demonstrate the effectiveness of the proposed method. Finally, some concerns about this method, such as model selection, estimation errors and application potential, are thoroughly analyzed and clarified.

eess.SY↗

arXiv2026

Improving the Reasoning of Multi-Image Grounding in MLLMs via Reinforcement Learning

Multimodal Large Language Models (MLLMs) perform well in single-image visual grounding but struggle with real-world tasks that demand cross-image reasoning and multi-modal instructions. To address this, we adopt a reinforcement learning (RL) based post-training strategy for MLLMs in multi-image grounding tasks. We first synthesize high-quality chain-of-thought (CoT) data for cold-start initialization, followed by supervised fine-tuning (SFT) using low-rank adaptation (LoRA). Subsequently, we apply rejection sampling with the merged SFT model to curate reliable RL data and use rule-based RL to guide the model toward optimal reasoning paths. Extensive experiments demonstrate the effectiveness of our approach, achieving +9.04% on MIG-Bench and +4.41% on average across seven out-of-domain benchmarks.

cs.CV↗

arXiv2026

HAWK: Head Importance-Aware Visual Token Pruning in Multimodal Models

In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications. Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens. Existing research usually assumes that all attention heads contribute equally to the visual interpretation. However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing. In light of this observation, we propose HAWK, a head importance-aware visual token pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens. By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones. The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs. Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0% of the original accuracy after pruning 80.2% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4% of the original and further decreases GPU memory usage across the tested models. The code is available at https://github.com/peppery77/HAWK.git.

cs.CV↗

arXiv2026

An Algorithm for Fast Assembling Large-Scale Defect-Free Atom Arrays

It is widely believed that tens of thousands of physical qubits are needed to build a practically useful quantum computer. Atom arrays formed by optical tweezers are among the most promising platforms for achieving this goal, owing to the excellent scalability and mobility of atomic qubits. However, assembling a defect-free atom array with ~ 10^4 qubits remains algorithmically challenging, alongside other hardware limitations. This is due to the computationally hard path-planning problems and the time-consuming generation of suffciently smooth trajectories for optical tweezer potentials by spatial light modulators (SLM). Here, we present a unified framework comprising two innovative components to fully address these algorithmic challenges: (1) a path-planning module that employs a supervised learning approach using a graph neural network combined with a modified auction decoder, and (2) a potential-generation module called the phase and profile-aware Weighted Gerchberg-Saxton algorithm. The inference time for the first module is nearly a size-independent constant overhead of ~ 5 ms, and the second module generates a potential frame with about 0.5 ms, a timescale shorter than the current commercial SLM refresh time. Altogether, our algorithm enables the assembly of an atom array with 10^4 qubits on a timescale much shorter than the typical vacuum lifetime of the trapped atoms.

cond-mat.quant-gas↗