arXiv subjects

Zheng Zhang

Publications and source records attributed to Zheng Zhang.

At least 19 recordsLinked to original sources

arXiv2026

PASA-KV: Attacking Action-Facing Visual Routes in World Action Models

World action models (WAMs) increasingly couple internal visual or world representations with action generation, making their world-to-action pathways a critical yet poorly understood source of robustness. We investigate when perturbations to these internal representations actually become behaviorally consequential, and uncover a pronounced mismatch between representation perturbability and closed-loop vulnerability: attack effectiveness varies sharply across visual K/V routes, while larger representation displacement does not reliably imply stronger control failure. Our analysis attributes this mismatch to the action-conditioned readout, where keys determine where action queries read and values determine what content they receive. This motivates an action-facing view of vulnerability, in which the effect of a perturbation depends on how it is routed into action generation rather than on representation-space magnitude alone. We instantiate this view with Tail-AV, a short-probe signal for identifying vulnerable routes, and operationalize it in PASA-KV, which selects action-facing K/V targets before training a fresh task-shared attack. Experiments across multiple WAM systems and a physical robot validate readout-informed target selection, demonstrate strong closed-loop attacks, and reveal substantial system-level differences in resistance.

cs.RO↗

arXiv2026

Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models

Vision-language models (VLMs) exhibit strong zero-shot generalization but remain vulnerable to adversarial attacks. Existing adversarial fine-tuning methods often use pre-trained model guidance to preserve image--text alignment while improving robustness. Our analysis of visual representation drift reveals complementary limitations of two representative forms of guidance. Matching visual features provides broad coverage by constraining all directions equally, but lacks semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities. These limitations can hinder preservation of pre-trained image--text relationships and degrade zero-shot performance. To address these issues, we propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings. Specifically, TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint. Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.

cs.CV↗

arXiv2026

Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization

Zeroth-order (ZO) optimization provides a gradient-free alternative to first-order (FO) methods by estimating gradients via finite differences of function evaluations, and has recently emerged as a memory-efficient paradigm for fine-tuning large-scale models by avoiding backpropagation. However, ZO optimization has a fundamental tension between accuracy and query efficiency. In this work, we show that ZO optimization can be substantially improved by unifying two complementary principles: (i) a projection-based subspace view that reduces gradient estimation variance by exploiting the intrinsic low-rank structure of model updates, and (ii) Muon-style spectral optimization that applies gradient orthogonalization to extract informative spectral structure from noisy ZO gradients. These findings form a unified framework of subspace gradient orthogonalization, which we instantiate in a new method, ZO-Muon, admitting a natural interpretation as a low-rank Muon optimizer in the ZO setting. Extensive experiments on large language models (LLMs) and vision transformers (ViTs) demonstrate that ZO-Muon significantly accelerates convergence and achieves a win-win improvement in accuracy and query/runtime efficiency. Notably, compared to the popular MeZO baseline, ZO-Muon requires only 24.7% of the queries to reach the same SST-2 performance for LLM fine-tuning, and improves accuracy by 25.1% on ViT-L fine-tuning on CIFAR-100.

cs.LG↗

arXiv2026

sft-wick: A formalism and package for Feynman-diagram expansion and evaluation in stochastic field theories

When stochastic field dynamics are cast into a path-integral formulation, perturbation theory becomes systematic but the resulting expansion quickly grows combinatorially large. The setting targeted here includes fields with several components and several dimensions, matrix propagators, tensor-valued couplings, and non-Gaussian driving noise specified by arbitrary n-point cumulants. Wick pairings grow factorially, and component indices must be routed through the tensor-valued vertices. The useful output is not a raw contraction list, but a diagram table: one entry per topology, with multiplicities, coupling sums, signs, and causal constraints resolved. We present sft-wick, an open-source Python package that constructs these diagram tables and computes their integrals numerically. Given an action and an observable, it enumerates topologically distinct Feynman diagrams, derives their algebraic coefficients, and evaluates the resulting diagram integrals from user-supplied response and cumulant functions. The core algorithm enumerates spatial topologies before routing component indices, avoiding contraction-by-contraction Wick expansion. Response-field constraints, including vanishing response-response contractions, the Ito prescription, and the absence of causal response loops, are enforced during enumeration. Predictions are validated against direct Langevin simulations in two worked examples, with Gaussian and with non-Gaussian driving noise, and agree with the simulations within stated error budgets.

physics.comp-ph↗

arXiv2026

p-DCIM: A 28nm Probabilistic SRAM DCIM based Ising Machine for Near-Memory Context Selection in LLM-RAG Pipeline

We present a fabricated SRAM DCIM Ising machine that performs set-level context selection for large language model (LLM) with retrieval-augmented generation (RAG). In RAG, the evidence-selection stage is increasingly delegated to repeated language-model reranking, which adds substantial energy to the inference pipeline. We reformulate evidence selection as a quadratic unconstrained binary optimization (QUBO) problem and solve it on a dual-mode digital compute-in-memory (DCIM) macro: the same multiply-and-accumulate datapath computes the embedding-similarity scores that populate the QUBO coefficients and evaluates the per-spin Hamiltonian differences that drive the anneal. The annealing temperature is set by a digital plane-count control along with the SRAM pseudo-read random bit flips, giving graded cooling entirely from on-chip digital configuration. The macro is taped-out in a 28 nm process [0.16 mm2, 0.41 mW] and is proposed to be integrated as part of the logic base-die in hybrid high-bandwidth memory (HBM) and high-bandwidth flash (HBF) stacks in a GPU package, where only the selected passage identifiers return to the host. Replacing the language-model selection pass with the near-memory Ising solver yields a macro normalized selection-energy estimate more than six orders of magnitude lower than the LLM-based baseline and reduces the evaluated total energy per query by 67%, while maintaining comparable token-level F1 on the evaluated examples.

cs.AR↗

arXiv2026

Beyond Single-Attribute Auditing: A Configurable Multi-Attribute Benchmark for Intersectional Fairness in LLMs

Large language models (LLMs) are increasingly evaluated for fairness, yet most existing benchmarks audit one sensitive attribute at a time. This single-attribute paradigm can miss intersectional harms, where multiple attributes interact non-additively and produce failures that are invisible under marginal evaluation. We introduce ComFair, a configurable benchmark that converts existing fairness data into 479,066 controlled Dual- and Multi-attribute instances and pairs every augmented instance with its task-equivalent source. Across Recognition, Selection, Continuation, and Conversation, this pairing supports a unified audit of performance or bias-score shifts, how often added attributes introduce new errors or mask existing ones, and model-ranking stability. Experiments on ten LLMs show task- and subgroup-dependent sensitivity: discriminative tasks expose both transition directions even when aggregate changes are modest, generative tasks exhibit larger and directionally asymmetric shifts, and rankings vary with attribute composition. ComFair therefore provides a scalable controlled sensitivity audit---rather than a model of the natural joint distribution of identities---for testing the robustness of fairness-benchmark conclusions beyond single-attribute settings.

cs.AI↗

arXiv2026

Agents' Last Exam

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long horizon, economically valuable, real world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 sub fields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is below 1%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP relevant impact.

cs.AI↗

arXiv2026

WorldFailBench: Do World Models Understand Policy Failures?

Robot policy evaluation informs deployment decisions and policy improvement, but physical testing and simulator preparation require substantial effort. World models offer a complementary route through closed-loop interaction in generated environments. Recent benchmarks have demonstrated their potential as policy evaluators through outcome-based evaluation (i.e., assessing task success or failure). However, such evaluation does not establish whether world models faithfully simulate how policies fail. Misrepresenting failures can lead to incorrect diagnoses of policy weaknesses and make policies appear ready for deployment even when small scene changes cause them to fail. We therefore ask: do world models faithfully simulate how and at which stages policies fail, and under what conditions they transition from success to failure? To answer this question, we introduce WorldFailBench with three evaluation tasks. Failure reason tests whether models reproduce the observable errors policies make. Failure stage assesses whether failures concentrate at the same stages of full-task execution. Failure boundary examines whether models reproduce the conditions under which policies transition from success to failure under controlled scene changes in simulation. The benchmark covers manipulation tasks in the LIBERO and RoboTwin simulation environments and real-world manipulation tasks performed by the G1 humanoid robot. Experiments across seven world models reveal substantial gaps in their understanding of policy failures, providing guidance for developing more reliable world models for policy evaluation.

cs.RO↗

arXiv2026

Linear Self-Symmetries of Inverse-Based S-boxes: Generic Rigidity and Applications

Finite-field inversion is a basic nonlinear building block in the design of S-boxes for symmetric cryptography. In inverse-based S-boxes, finite-field inversion is often combined with linear or affine layers that may preserve or destroy algebraic symmetries of the inverse function. We study how often nontrivial linear self-symmetries remain as the outer invertible linear transformation varies over $GL_n(\mathbb F_2)$. Using the classification of two-sided linear equivalences of inversion, we reduce the problem to semilinear conjugacy and centralizer estimates. We obtain matching upper and lower bounds for the probability that a nontrivial linear self-symmetry remains, and these bounds determine its leading quadratic exponent. This exponent is governed by the smallest prime divisor of the extension degree. Thus the decay rate depends on the arithmetic of the field extension and not only on its dimension. As cryptographic applications, we provide a practical rigidity criterion, prove that the full AES S-box has no nontrivial linear self-symmetry, and illustrate the broader relevance of the framework through ARIA and Vision.

cs.CR↗

arXiv2026

Infinitesimal Torelli problems for special Gushel-Mukai and related Fano threefolds: Hodge theoretical and categorical perspectives

A hyperelliptic curve is the classical example of a variety for which the global Torelli theorem holds while the differential of the period map has a kernel, forced by the extra involution. We study the analogous phenomenon for Fano threefolds that are double covers of rigid Fano threefolds branched along a K3 surface, mainly special Gushel--Mukai threefolds. We prove that the part of the infinitesimal period map invariant under the covering involution is injective, so that the kernel is exactly the three-dimensional space of anti-invariant deformations, one more than for ordinary Gushel--Mukai threefolds. We give a Hodge-theoretic proof, which reduces to a twisted infinitesimal Torelli theorem for the branch K3 surface, and a categorical proof, based on Kuznetsov's normal Hochschild cohomology, which also applies in genus eight and identifies the kernel with the tangent space to a distinguished point of a Bridgeland moduli space. The two descriptions of the kernel are related directly through the Plücker net and the conic of special lines of the quintic del Pezzo threefold. Finally, ordinary Verra threefolds satisfy the infinitesimal Torelli theorem, while special ones have a one-dimensional kernel; in both cases the kernel agrees with that of the categorical period map of a Clifford component, and it is the tangent space to the deformations of a canonical spinor bundle, which is rigid in the ordinary case and obstructed at second order in the special case.

math.AG↗

arXiv2026

T-MUON: Tensorizing Muon via t-SVD for Large Language Model Pre-training

Muon has recently emerged as a strong optimizer for large language model pretraining by exploiting matrix structure through polar orthogonalization. However, it processes each parameter matrix independently, whereas recent tensor-aware extensions of Muon jointly update groups of eligible matrix parameters through matrix unfoldings. In contrast, we propose T-MUON, a t-SVD-based generalization that groups structurally related Transformer gradients into third-order tensors and defines the corresponding Muon-style update directly through the $M$-transform. We show that the exact T-MUON update is a steepest-descent direction under the $\star_M$-spectral norm and establish a convergence guarantee. Across GPT-2-style and LLaMA-style models from 51M to 1.3B parameters trained on FineWeb and FineWeb-Edu, head-wise T-MUON with the identity transform consistently achieves lower validation perplexity than AdamW, Muon, and variants using DFT, DCT, or DST transforms. These gains persist in additional experiments with longer context lengths and larger token budgets, although the preferred tensorization mode varies across settings.

cs.LG↗

arXiv2026

Rep2Skill: Representation-Guided Skill Self-Evolution for LLM Agents

Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models their internal model representation trajectories to localize turns that deviate from successful execution dynamics, and it further interprets these signals alongside the execution contexts as actionable textual feedback for targeted skill revision. Experiments on two agent environments with two open-source LLMs show that Rep2Skill consistently outperforms text-only approaches in the self-evolution setting, where the same LLM serves as both executor and optimizer without a stronger external model. This establishes a promising direction moving agent self-improvement beyond text-only reflection.

cs.AI↗

arXiv2026

PatientWM: Towards Generalizable Clinical World Models via Unified Patient States

Clinical decisions depend on anticipating how a patient's condition will evolve as treatment is adjusted and new evidence becomes available. Existing medical world models capture patient-state evolution, but often define the evolving state around a particular modality or task, limiting reuse as inputs or prediction targets change. These models also retain history through observation sequences or recurrent or latent summaries rather than exposing preceding patient states and clinical actions as an ordered transition context. This distinction matters because similar current measurements may follow different recent changes, treatment exposures, and observed responses. To tackle these limitations, we propose PatientWM, a clinical world-model framework built around a unified patient state and sequential latent history. Availability-aware encoders integrate observations available at each decision time into the patient state, while typed decoders produce task-specific clinical readouts without defining the state around any single target. PatientWM further conditions each next-state prediction on the current state and supplied action together with a bounded, ordered history of preceding latent patient states and clinical actions, preserving recent treatment and response context. Experiments across MIMIC-III, MIMIC-IV, and three external clinical datasets show superior performance in clinical prediction, multi-step forecasting, and action-planning evaluation. Transfer experiments further demonstrate the generalizability of PatientWM's patient state. Together, these results establish its ability to forecast clinical evolution and support planning.

cs.AI↗

arXiv2026

Enhancing Quasi-Two-Dimensional Electronic Structure in Bi2O2Se via Chlorine Substitution

Bismuth oxyselenide (Bi2O2Se) has emerged as a leading semiconductor candidate for next-generation electronics and quantum phenomena research. However, its strong ionic interlayer bonding of Se layers fundamentally restricts the full potential of this otherwise promising quasi-two-dimensional (2D) material. In this work, we demonstrate chlorine substitution as an effective strategy to introduce van der Waals gaps and modify the electronic structure of the system. Through comprehensive transport and angle-resolved photoemission spectroscopy (ARPES) measurements on Bi6O6Se2Cl2 single crystal, an optimal Cl-doping compound, we show that the key electronic features, including indirect band gap, high electron mobility, and carrier effective mass, are well preserved, while the 2D character of the system is sufficiently enhanced. First-principles calculations attribute this selective tunability to the unique spatial separation between conductive Bi-O layers and dopant Se/Cl layers, which enables independent optimization of interlayer coupling and intralayer transport. Remarkably, our theoretical calculations combined with experimental observations project a broad bandgap engineering across a wide range (0.8 ~ 3.5 eV) within the Bi2O2Se system. Our findings not only expand the application potential of Bi2O2Se-based materials in optoelectronics, nanoelectronics, and fundamental research but also highlight spatial decoupling as a viable strategy for modifying detailed electronic structure of quasi-2D quantum systems.

cond-mat.mtrl-sci↗

arXiv2026

Evolving Creatives with Clicks: CTR-Driven Advertising image Editing via Attribute Disentanglement and Preference Alignment

The optimization objective for e-commerce advertising creatives has shifted from aesthetic quality to online business metrics like click-through rate (CTR). Recent methods generate appealing ad images from product images by constructing CTR-based reward signals and refining prompt generation via Reinforcement Learning. Such paradigms equally penalize all attributes within an underperforming image, even though some are actually beneficial to clicks. In this work, we propose Evolving Creatives with Clicks (ECC), which reformulates creative optimization as fine-grained editing rather than generation from scratch. ECC selectively refines attributes that hurt CTR while preserving those that contribute positively. Specifically, faced with a rich attribute space across massive images, ECC first identifies 28 attributes significantly correlated with CTR via ANOVA and Random Forest regression. Based on attribute contrasts between images with disparate CTRs, we instruction-tune a prompt model to generate editing suggestions, then refine it with real click preferences since reference-derived suggestions are not uniformly beneficial. Furthermore, we introduce Consistent Preference Optimization (CPO) as ECC's preference optimization strategy, which enforces consistency across semantically equivalent shuffled suggestions to mitigate DPO overfitting and improve preference generalization. Extensive experiments demonstrate that ECC significantly improves the CTR of ad images, outperforming state-of-the-art models by a substantial margin. The complete data pipeline, including attribute extraction and analysis, the resulting dataset, and source code will be released upon acceptance.

cs.CV↗

arXiv2026

Statistical Field Theory for Weak Gravitational Lensing

Standard weak-lensing calculations treat lensing as a linear remapping of the matter field along the line of sight. We instead formulate lensing as a stochastic field theory for the Sachs optical scalars, driven by random Ricci-focusing and Weyl-shearing fields. With convergence and shear expressed linearly in the Sachs-scalar fluctuations, the resulting path integral generates a diagrammatic expansion for their $n$-point correlation functions, organised into linear response, nonlinear propagation, and driving-field cumulants. The conventional calculation emerges as the lowest-order, linear-propagation limit. Beyond it, nonlinear Sachs evolution couples to driving-field non-Gaussianity, mixing the matter cumulant hierarchy into the lensing hierarchy. A selection rule governs the couplings: an $n$-point observable receives a direct contribution from the $n$-point driving-field cumulant, and its leading hierarchy-mixing correction from the $(n+1)$-point cumulant via one nonlinear Sachs interaction, with higher cumulants entering only at higher order. The two-point function, for instance, receives corrections from three-point cumulants of Ricci focusing and Weyl shearing through nonlinear Sachs evolution. Rather than a restrictive approximation scheme, the formalism is a paradigm shift: a unified framework naturally accommodating path corrections, higher-order matter statistics, stochasticity, and small-scale effects.

astro-ph.CO↗

arXiv2026

LoRi: Low-Rank Distillation for Implicit Reasoning

Implicit chain-of-thought (iCoT) methods aim to internalize reasoning in large language models, but often underperform explicit CoT prompting. We empirically find that hidden-state reasoning trajectories exhibit low-rank structure. Motivated by this observation, we propose a low-rank distillation framework that transfers reasoning by aligning teacher and student trajectories in a shared low-rank tensor subspace using first- and second-order statistics. The resulting formulation captures the global structure of reasoning while supporting a compact latent reasoning process. We evaluate the method across multiple model families, including LLaMA and Qwen, at different scales on mathematical reasoning benchmarks. Our approach consistently improves performance, especially on challenging multi-step tasks, approaching explicit CoT accuracy and outperforming prior iCoT distillation methods.

cs.CL↗

arXiv2026

AdSpark: A Large-Scale Dataset and Benchmark for Product-Centric Advertisement Video Generation

Product-centric advertisement video generation aims to create promotional videos that preserve fine-grained product identity while presenting selling points through coherent multi-shot narratives. However, this emerging task remains underexplored due to the lack of large-scale advertisement-specific datasets and comprehensive evaluation frameworks. To address this gap, we introduce \textbf{AdSpark}, a large-scale dataset and benchmark for product-centric advertisement video generation, based on data from a major e-commerce platform. \textit{AdSpark-300K} contains approximately 300K reference image--prompt--video triplets, comprising a real-world subset and a synthetic subset. Each sample provides structured advertisement annotations, including product identity annotations, selling-point descriptions, creative plans, and aligned audio scripts, enabling models to learn product preservation and advertisement-oriented visual storytelling. We further propose \textit{AdSpark-Bench}, a diagnostic benchmark that evaluates generated advertisements across six dimensions, including visual quality, product fidelity, instruction adherence, temporal coherence, audio alignment, and advertisement effectiveness. Based on AdSpark-Bench, we evaluate representative models, revealing key challenges in product preservation, multi-shot storytelling, and selling-point visualization. Experiments with AdSpark-300K-finetuned models further validate the effectiveness of our dataset. AdSpark provides a unified dataset and benchmark for future research, and we will release the dataset upon acceptance.

cs.CV↗