Searcharxiv⌕ Search

arXiv subjects

Ziqi Zhao

Publications and source records attributed to Ziqi Zhao.

At least 19 recordsLinked to original sources
arXiv2026

Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges

Safety benchmarks for large language models often assess the risk of a user query, although the outcome of question answering depends on whether the response violates a policy. This distinction is critical in Chinese harmful-content evaluation, where linguistic variation and adversarial transformations can obscure risky intent. We introduce C-SafeQA, a policy-grounded benchmark for response-level Chinese safety evaluation. It comprises 538 base queries and 8,877 adversarial queries answered by four full-model LLM deployments, yielding 37,660 query-response records labeled safe, unsafe, or disputed. Reference labels are generated through agreement-aware multi-model adjudication and blind audits of stratified subsets by three safety experts. C-SafeQA supports both evaluation of target-model safety and auditing of seven automated safety judges against shared reference labels. Unsafe-response rates range from 0.93% to 3.35% on base queries and from 11.68% to 30.05% on adversarial queries. On the adversarial subset, judges show substantial trade-offs between unsafe-response recall and risk-query-conditioned safe-response false positive rate, and no judge dominates all metrics. Both acrostic transformations reduce unsafe recall for all seven judges, revealing mechanism-specific evaluator weaknesses. Dataset records, metadata, verification code, and judge scripts are publicly released to support recomputation, while benchmark construction, target-response generation, and private adjudication remain outside the release boundary.

cs.CR↗
arXiv2026

Matched Queries for Curvature and Density at Branching Junctions

At a junction, a score field can reveal weighted tangent rays, yet these first-order quantities do not determine how individual branches bend or how their densities change away from the center. Recovering this missing information is necessary for describing local continuation beyond a single point, but finite observations must separate branchwise second-order effects while allowing error in the estimated center. We address this inverse problem using matched score queries at noise scales $\sigma$ and $\lambda\sigma$. For a finite union of $C^{2,\alpha}$ half-branches in $\mathbb{R}^D$, the normalized score has the expansion $F_\sigma=F_0+\sigma G+O(\sigma^{1+\alpha})$. Matched subtraction cancels the tangent contribution and exposes $G$, which depends linearly on branchwise curvature and log-density slope. Given tangent directions and weights on distinct rays, $G$ uniquely identifies all $sD$ branch parameters, and $sD$ scalar component observations are necessary. An $O(\sigma^2)$ center error introduces $D$ translation modes, leading to $(s+1)D$ observations under full-rank calibration, except for a translation-invariant full line. We also establish a perturbation bound and a conditional kernel-density-estimation rate. Experiments reproduce the predicted population and $N^{-1/5}$ trends and remain full rank up to $D=20$ with 16 supplied branches. In end-to-end tests for $D=3$--$5$, a known-count first-order frontend yields full rank in all 135 population systems and a median relative jet error of 0.132. With strong first-order error, matched responses reduce median parameter error by a factor of 49.4 relative to naive tangent subtraction.

stat.ML↗
arXiv2026

Recovering Weighted Tangent Geometry from a Single-Scale Score Field

Near a smooth data manifold, one tangent space summarizes local geometry. At a branch point, the corresponding first-order object is instead a measure over tangent directions, whose normalized masses record the local share of each branch under the chosen data measure. We ask whether a score field at one noise level determines this weighted tangent geometry when the branch center and homogeneity degree $d$ are unknown. In this tangent-measure model, $d$ is the local measure dimension. Gaussian smoothing of a homogeneous tangent measure satisfies an Ornstein--Uhlenbeck eigenfunction equation. Its weak form turns score values---without score derivatives---into a linear system for the center and homogeneity degree, with an explicit rank condition and perturbation bound. After this calibration, the tangential score on one sphere is the spherical log-gradient of a scalar Gaussian--cone transform. Integration recovers that transform up to scale, and all its spherical-harmonic multipliers are positive. Thus one exact shell identifies the normalized angular measure in every ambient dimension $D\geq2$. For at most $K$ positive rays, moments through degree $2K-1$ constructively recover count, directions, and weights in arbitrary dimension. Any fixed observation scheme needs at least $KD-1$ scalar tangential components. In the plane, degree $K$ is both sufficient and necessary, and we give quantitative finite-query certificates. For finite planar $C^{1,β}$ branches with positive $C^{0,β}$ densities, we prove $O(σ^β)$ convergence from the finite-noise score to its tangent model. In controlled experiments, 50k-step training lowers validation normalized-score error across four geometries yet raises angular-moment error, separating ordinary score fit from geometry recovery.

stat.ML↗
arXiv2026

When More Generators Hurt: Shellsort on Full Product Grids

Shellsort repeatedly runs insertion sort with decreasing gaps, so its worst-case cost depends on the gap sequence. Pratt's $2^u3^v$ sequence, one of the few systematic constructions with a proven $O(n\log^2 n)$ bound, includes every product below $n$ of two base numbers, or generators. We ask whether adding more base numbers, and thus more intermediate gaps, can improve this full product grid. We show that it cannot when every product is retained and each base is at most a fixed power of the smallest. With $r$ independent bases (different exponent choices give different products) and $Θ(\log n)$ gaps, the best possible worst-case cost is $n\exp(Θ((\log n)^{1-1/r}))$. Thus two bases give the exponent $\sqrt{\log n}$, whereas three give $(\log n)^{2/3}$: more bases are worse. With a budget of $p$ gaps, matching bounds give the factor $\exp(Θ(\log n/p^{1/r}))$ beyond linear cost. The reason is simple. Few products force the smallest base $m$ to be large, and fullness makes $m$ the next-to-last gap. An input built from reversed blocks is already sorted for every earlier gap, forcing $Ω(nm)$ work in the final pass. Powers of distinct primes give a matching construction. For arbitrary gaps, we count current-gap multiples that earlier gaps cannot form. This gives upper and lower bounds for individual passes. A Fourier argument gives necessary conditions for small total cost, while short nonnegative sums give sufficient conditions. In both settings, useful distances must be available before they are needed.

cs.CC↗
arXiv2026

Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice

Vertical Federated Learning (VFL) enables organizations holding complementary features of shared entities to collaborate and train models. In this setting, the initiator can withhold information about the learning task, while other contributors participate without exposing their local datasets, creating an asymmetric information structure aligned with growing privacy demands. However, this asymmetry is a double-edged sword. Among various threats, backdoor attacks are particularly concerning because VFL not only enables malicious contributors to poison the model during training, but also allows them to activate the backdoor at inference time to manipulate predictions. Although prior work has reported near-perfect attack success rates and proposed effective defenses, we find that most findings fail to hold under realistic conditions, exposing a fundamental gap between research and practice. In this paper, we present a systematic, practice-oriented study of backdoor vulnerabilities in VFL, revealing this gap in both methodological design and evaluation practices. We show that existing approaches overlook key practical constraints and therefore rely on unrealistic prior knowledge. Furthermore, these limitations have remained hidden due to poorly designed evaluation practices in the literature. To bridge this gap, we redefine threat models under realistic constraints, propose practical backdoor workflows, and introduce BVBench, a backdoor-centric benchmark that enables fair, practical, and comprehensive evaluation, preloaded with state-of-the-art baselines. BVBench provides strong evidence of the fragility of the current understanding of VFL backdoor risks and establishes a foundation for steering research toward uncovering practical vulnerabilities and developing more meaningful defenses.

cs.LG↗
arXiv2026

RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33\% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26\% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

cs.SE↗
arXiv2026

Formation of Extremely Low-mass White Dwarf Binaries Undergoing Enhanced Angular Momentum Loss

Extremely low-mass white dwarfs (ELM WDs) are helium (He) WDs with masses below $\sim 0.3\ M_{\odot}$, mainly formed through binary interaction. ELM WD binaries typically are formed from two channels, namely the stable Roche lobe overflow (RLOF) channel and the common envelope ejection channel. For ELM WD binaries produced from RLOF channel, the ELM WD mass has a strong correlation with the orbital period, i.e., the so-called WD mass-orbital period relation. However, the observations in the ELM Survey show that the orbital periods of ELM WD binaries from the RLOF channel are typically shorter than the theoretically predicted values. Extra angular momentum loss (AML) may be needed to explain such a phenomenon. In this work, we assumed that part of the transferred mass from the donor is lost at the outer Lagrangian point and simulated the formation of ELM WD binaries. Enhanced AML enables more mass to be lost during thermal-timescale mass transfer, thereby affecting nuclear burning in the transfer phase and producing ELM WDs with distinct internal structures. These structural differences alter the (pre-)He WD mass-radius relation at the end of mass transfer, which in turn shifts the WD mass-orbital period relation downward. These adjustments enable our model to successfully reproduce the majority of observed systems from the relevant survey projects.

astro-ph.SR↗
arXiv2026

Symmetry of the critical current function in superconducting nanodevices

We study a variety of nano-scale superconducting devices containing more than one weak link. Even in devices with multiple weak links the IB symmetry is usually obeyed, which means that if we reverse both the bias current direction and the magnetic field direction at the same time, the superconducting response, namely the critical current, remains unchanged. We also provide a detailed analysis of the situations in which such symmetry is violated.

cond-mat.supr-con↗
arXiv2026

ROSD: Reflective On-Policy Self-Distillation for Language Model Reasoning across Domains

On-policy self-distillation (OPSD) improves the reasoning performance of large language models (LLMs) by providing dense token-level supervision for on-policy rollouts. However, existing OPSD methods often yield limited gains on in-domain reasoning and generalize poorly to out-of-domain problems. We identify two key causes: conditioning the self-teacher on a verified solution encourages imitation of training-domain reference trajectories rather than error-specific correction, and applying distillation to the full response can overwrite valid reasoning prefixes and reinforce overfitting. We propose Reflective On-policy Self-Distillation (ROSD), a framework that turns reference-solution imitation into targeted reasoning correction through reflection-guided, error-localized distillation. For each rollout, ROSD uses a self-reflector to extract a corrective idea and locate the first erroneous span. The corrective idea guides the self-teacher toward targeted supervision, while the localized error span restricts distillation to where correction is needed. This design corrects flawed reasoning while preserving valid prefixes. Experiments on multiple in-domain and out-of-domain reasoning benchmarks show that ROSD yields stronger in-domain reasoning performance overall and substantially better out-of-domain generalization than standard OPSD. Code is available at https://github.com/ZiqiZhao1/ROSD.

cs.CL↗
arXiv2026

ClawMark: A Living-World Benchmark for Multi-Turn, Multi-Day, Multimodal Coworker Agents

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

cs.CV↗
arXiv2026

Reinforced Efficient Reasoning via Semantically Diverse Exploration

Reinforcement learning with verifiable rewards (RLVR) has proven effective in enhancing the reasoning of large language models (LLMs). Monte Carlo Tree Search (MCTS)-based extensions improve upon vanilla RLVR (e.g., GRPO) by providing tree-based reasoning rollouts that enable fine-grained and segment-level credit assignment. However, existing methods still suffer from limited exploration diversity and inefficient reasoning. To address the above challenges, we propose reinforced efficient reasoning via semantically diverse explorations, i.e., ROSE, for LLMs. To encourage more diverse reasoning exploration, our method incorporates a semantic-entropy-based branching strategy and an $\varepsilon$-exploration mechanism. The former operates on already sampled reasoning rollouts to capture semantic uncertainty and select branching points with high semantic divergence to generate new successive reasoning paths, whereas the latter stochastically initiates reasoning rollouts from the root, preventing the search process from becoming overly local. To improve efficiency, we design a length-aware segment-level advantage estimator that rewards concise and correct reasoning while penalizing unnecessarily long reasoning chains. Extensive experiments on various mathematical reasoning benchmarks with Qwen and Llama models validate the effectiveness and efficiency of ROSE. Codes are available at https://github.com/ZiqiZhao1/ROSE-rl.

cs.AI↗
arXiv2026

Gym-V: A Unified Vision Environment System for Agentic Vision Research

As agentic systems increasingly rely on reinforcement learning from verifiable rewards, standardized ``gym'' infrastructure has become essential for rapid iteration, reproducibility, and fair comparison. Vision agents lack such infrastructure, limiting systematic study of what drives their learning and where current models fall short. We introduce \textbf{Gym-V}, a unified platform of 179 procedurally generated visual environments across 10 domains with controllable difficulty, enabling controlled experiments that were previously infeasible across fragmented toolkits. Using it, we find that observation scaffolding is more decisive for training success than the choice of RL algorithm, with captions and game rules determining whether learning succeeds at all. Cross-domain transfer experiments further show that training on diverse task categories generalizes broadly while narrow training can cause negative transfer, with multi-turn interaction amplifying all of these effects. Gym-V is released as a convenient foundation for training environments and evaluation toolkits, aiming to accelerate future research on agentic VLMs.

cs.CV↗
arXiv2026

TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition

Table recognition (TR) aims to transform table images into semi-structured representations such as HTML or Markdown. As a core component of document parsing, TR has long relied on supervised learning, with recent efforts dominated by fine-tuning vision-language models (VLMs) using labeled data. While VLMs have brought TR to the next level, pushing performance further demands large-scale labeled data that is costly to obtain. Consequently, although proprietary models have continuously pushed the performance boundary, open-source models, often trained with limited resources and, in practice, the only viable option for many due to privacy regulations, still lag far behind. To bridge this gap, we introduce TRivia, a self-supervised fine-tuning method that enables pretrained VLMs to learn TR directly from unlabeled table images in the wild. Built upon Group Relative Policy Optimization, TRivia automatically identifies unlabeled samples that most effectively facilitate learning and eliminates the need for human annotations through a question-answering-based reward mechanism. An attention-guided module generates diverse questions for each table image, and the ability to interpret the recognition results and answer them correctly provides feedback to optimize the TR model. This closed-loop process allows the TR model to autonomously learn to recognize, structure, and reason over tables without labeled data. Leveraging this pipeline, we present TRivia-3B, an open-sourced, compact, and state-of-the-art TR model that surpasses existing systems (e.g., Gemini 2.5 Pro, MinerU2.5) on three popular benchmarks. Model and code are released at: https://github.com/HKU-TASR/TRivia

cs.CV↗
arXiv2026

LoR-LUT: Learning Compact 3D Lookup Tables via Low-Rank Residuals

We present LoR-LUT, a unified low-rank formulation for compact and interpretable 3D lookup table (LUT) generation. Unlike conventional 3D-LUT-based techniques that rely on fusion of basis LUTs, which are usually dense tensors, our unified approach extends the current framework by jointly using residual corrections, which are in fact low-rank tensors, together with a set of basis LUTs. The approach described here improves the existing perceptual quality of an image, which is primarily due to the technique's novel use of residual corrections. At the same time, we achieve the same level of trilinear interpolation complexity, using a significantly smaller number of network, residual corrections, and LUT parameters. The experimental results obtained from LoR-LUT, which is trained on the MIT-Adobe FiveK dataset, reproduce expert-level retouching characteristics with high perceptual fidelity and a sub-megabyte model size. Furthermore, we introduce an interactive visualization tool, termed LoR-LUT Viewer, which transforms an input image into the LUT-adjusted output image, via a number of slidebars that control different parameters. The tool provides an effective way to enhance interpretability and user confidence in the visual results. Overall, our proposed formulation offers a compact, interpretable, and efficient direction for future LUT-based image enhancement and style transfer.

cs.CV↗
arXiv2026

Curriculum Approximate Unlearning for Session-based Recommendation

Approximate unlearning for session-based recommendation refers to eliminating the influence of specific training samples from the recommender without retraining of (sub-)models. Gradient ascent (GA) is a representative method to conduct approximate unlearning. However, there still exist dual challenges to apply GA for session-based recommendation. On the one hand, naive applying of GA could lead to degradation of recommendation performance. On the other hand, existing studies fail to consider the ordering of unlearning samples when simultaneously processing multiple unlearning requests, leading to sub-optimal recommendation performance and unlearning effect. To address the above challenges, we introduce CAU, a curriculum approximate unlearning framework tailored to session-based recommendation. CAU handles the unlearning task with a GA term on unlearning samples. Specifically, to address the first challenge, CAU formulates the overall optimization task as a multi-objective optimization problem, where the GA term for unlearning samples is combined with retaining terms for preserving performance. The multi-objective optimization problem is solved through seeking the Pareto-Optimal solution, which achieves effective unlearning with trivial sacrifice on recommendation performance. To tackle the second challenge, CAU adopts a curriculum-based sequence to conduct unlearning on batches of unlearning samples. The key motivation is to perform unlearning from easy samples to harder ones. To this end, CAU first introduces two metrics to measure the unlearning difficulty, including gradient unlearning difficulty and embedding unlearning difficulty. Then, two strategies, hard-sampling and soft-sampling, are proposed to select unlearning samples according to difficulty scores.

cs.IR↗
arXiv2025

nuGPR: GPU-Accelerated Gaussian Process Regression with Iterative Algorithms and Low-Rank Approximations

Gaussian Process Regression (GPR) is an important type of supervised machine learning model with inherent uncertainty measure in its predictions. We propose a new framework, nuGPR, to address the well-known challenge of high computation cost associated with GPR training. Our framework includes several ideas from numerical linear algebra to reduce the amount of computation in key steps of GPR, and we combine them to establish an end-to-end training algorithm. Specifically, we leverage the preconditioned conjugate gradient method to accelerate the convergence of the linear solves required in GPR. We exploit clustering in the input data to identify block-diagonal structure of the covariance matrix and subsequently construct low-rank approximations of the off-diagonal blocks. These enhancements significantly reduce the time and space complexity of our computations. In addition, unlike other frameworks that rely on exact differentiation, we employ numerical gradients to optimize the hyperparameters of our GPR model, further reducing the training cost by eliminating the need for backpropagation. Lastly, we leverage the CUDA Toolkit to efficiently parallelize the training procedure on NVIDIA GPUs. As a result, nuGPR reduces total training time by up to 2x and peak memory consumption by up to 12x on various synthetic and real-world datasets when compared to the best existing GPU-based GPR implementation.

cs.LG↗
arXiv2025

Perfect Superconducting Diode and Supercurrent Range Controller

Diodes have a nonreciprocal voltage versus current relationship, produced by breaking the space and time reversal symmetry. However, developing high-end superconducting computers requires a superconducting analogue of the traditional semiconductor diode. Such a superconducting diode exhibits non-reciprocity, or a high asymmetry in its critical currents. We present a model of a perfect superconducting diode based on a superconducting quantum interference device made with multiple superconducting nanowires. The diode predicted by our model has a large positive critical current, while the negative critical current can be exactly zero. This 100\% diode efficiency ($η= 1$) remains stable against small changes of the magnetic field. Another important result is that under certain and quite broad conditions such devices can act as supercurrent range controllers. In such device a supercurrent can flow with zero voltage applied, but only if the supercurrent is contained in some narrow, adjustable range, which excludes zero current.

cond-mat.supr-con↗
arXiv2025

HealthiVert-GAN: A Novel Framework of Pseudo-Healthy Vertebral Image Synthesis for Interpretable Compression Fracture Grading

Osteoporotic vertebral compression fractures (OVCFs) are prevalent in the elderly population, typically assessed on computed tomography (CT) scans by evaluating vertebral height loss. This assessment helps determine the fracture's impact on spinal stability and the need for surgical intervention. However, the absence of pre-fracture CT scans and standardized vertebral references leads to measurement errors and inter-observer variability, while irregular compression patterns further challenge the precise grading of fracture severity. While deep learning methods have shown promise in aiding OVCFs screening, they often lack interpretability and sufficient sensitivity, limiting their clinical applicability. To address these challenges, we introduce a novel vertebra synthesis-height loss quantification-OVCFs grading framework. Our proposed model, HealthiVert-GAN, utilizes a coarse-to-fine synthesis network designed to generate pseudo-healthy vertebral images that simulate the pre-fracture state of fractured vertebrae. This model integrates three auxiliary modules that leverage the morphology and height information of adjacent healthy vertebrae to ensure anatomical consistency. Additionally, we introduce the Relative Height Loss of Vertebrae (RHLV) as a quantification metric, which divides each vertebra into three sections to measure height loss between pre-fracture and post-fracture states, followed by fracture severity classification using a Support Vector Machine (SVM). Our approach achieves state-of-the-art classification performance on both the Verse2019 dataset and in-house dataset, and it provides cross-sectional distribution maps of vertebral height loss. This practical tool enhances diagnostic accuracy in clinical settings and assisting in surgical decision-making.

eess.IV↗