arXiv subjects

Ning Zheng

Publications and source records attributed to Ning Zheng.

At least 19 recordsLinked to original sources

arXiv2026

Distillation as Probability Transport: Routed On-Policy Distillation

On-policy distillation (OPD) transfers teacher knowledge on student-generated trajectories, but efficient sampled objectives reduce the teacher distribution to scalar credit on individual tokens. Such credit indicates whether a token should gain or lose probability, yet leaves the corresponding redistribution unspecified. We recast OPD as teacher-guided probability transport and propose RouteOPD (Routed On-Policy Distillation), which decomposes local teacher--student disagreement into student-excess sources and teacher-deficit destinations and couples them into explicit transport pairs. RouteOPD optimizes pairwise log-odds toward jointly realizable targets obtained from a bounded teacher potential, while adapting the transport budget to the concentration of teacher demand. This formulation directs updates toward teacher-preferred destinations and controls their magnitude within a single transport operator. Experiments across four teacher--student settings and four mathematical-reasoning benchmarks demonstrate that RouteOPD consistently outperforms sampled reverse-KL OPD, with improvements accompanied by higher routing fidelity and lower background leakage. These results demonstrate the effectiveness of explicitly modeling probability transport in on-policy distillation.

cs.LG↗

arXiv2026

When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents

Self-evolving agents accumulate capability by distilling reusable skills from their execution trajectories, but we find this process is not monotonic: past a critical pool size, newly added skills degrade performance instead of improving it. We formalize this capability-contamination phase transition and trace it to a structural cause: once a defective skill enters the decision context, it becomes reference material for distilling later skills, forming cross-round contamination chains. We further show the contamination is structurally irreversible: removing a source skill after the fact cannot erase the flawed reasoning its descendants have already inherited, so post-hoc rollback recovers only a small fraction of the lost performance. This makes skill admission a pre-commit necessity rather than a post-hoc fix, and motivates Verifier-as-Gatekeeper (VaG): a progressive trust hierarchy whose three heterogeneous critics - structural validity, behavioral harmlessness, and semantic consistency - filter each skill individually, coupled with a marginal-gain subset selection that removes combinatorial contamination at the top tier before skills reach the runtime context. On Terminal-Bench 2, unconditional accumulation rises to a peak and then degrades, giving back most of its gains as the pool keeps growing, and post-hoc removal of the culprit skills recovers only a small part of the drop - the empirical signature of irreversibility. In contrast, VaG improves every round, reaching 72% pass@1 with a pool roughly 5x smaller, and its frozen skill pool transfers positively to four other backbones and a second benchmark without re-evolution. Ablations confirm the three critics are complementary and mutually non-substitutable, each intercepting a largely disjoint class of harmful skills.

cs.AI↗

arXiv2026

Evaluation of Large Language Models in Legal Applications: Challenges, Methods, and Future Directions

Large language models (LLMs) are being increasingly integrated into legal applications, including judicial decision support, legal practice assistance, and public-facing legal services. While LLMs show strong potential in handling legal knowledge and tasks, their deployment in real-world legal settings raises critical concerns beyond surface-level accuracy, involving the soundness of legal reasoning processes and trustworthy issues such as fairness and reliability. Systematic evaluation of LLM performance in legal tasks has therefore become essential for their responsible adoption. This survey identifies key challenges in evaluating LLMs for legal tasks grounded in real-world legal practice. We analyze the major difficulties involved in assessing LLM performance in the legal domain, including outcome correctness, reasoning reliability, and trustworthiness. Building on these challenges, we review and categorize existing evaluation methods and benchmarks according to their task design, datasets, and evaluation metrics. We further discuss the extent to which current approaches address these challenges, highlight their limitations, and outline future research directions toward more realistic, reliable, and legally grounded evaluation frameworks for LLMs in legal domains.

cs.CY↗

arXiv2025

Iterative Methods for Computing the Moore-Penrose Pseudoinverse of Quaternion Matrices, with Applications

We develop quaternion--native iterative methods for computing the Moore--Penrose (MP) pseudoinverse of quaternion matrices and analyze their convergence. Our starting point is a damped Newton--Schulz (NS) iteration tailored to noncommutativity: we enforce the appropriate left/right identities for rectangular inputs and prove convergence directly in $\mathbb{H}$ under a simple spectral scaling. We then derive higher--order (\emph{hyperpower}) NS schemes with exact residual recurrences that yield order-$p$ local convergence, together with factorizations that reduce the number of $s\times s$ quaternion products per iteration. Beyond NS, we introduce a randomized sketch--and--project method (RSP--Q), a hybrid RSP+NS scheme that interleaves inexpensive randomized projections with an exact hyperpower step, and a matrix--form conjugate gradient on the normal equations (CGNE--Q). All algorithms operate directly in $\mathbb{H}$ (no real or complex embeddings) and are matrix--free.

math.NA↗

arXiv2025

LLMs on Trial: Evaluating Judicial Fairness for Large Language Models

Large Language Models (LLMs) are increasingly used in high-stakes fields where their decisions impact rights and equity. However, LLMs' judicial fairness and implications for social justice remain underexplored. When LLMs act as judges, the ability to fairly resolve judicial issues is a prerequisite to ensure their trustworthiness. Based on theories of judicial fairness, we construct a comprehensive framework to measure LLM fairness, leading to a selection of 65 labels and 161 corresponding values. Applying this framework to the judicial system, we compile an extensive dataset, JudiFair, comprising 177,100 unique case facts. To achieve robust statistical inference, we develop three evaluation metrics, inconsistency, bias, and imbalanced inaccuracy, and introduce a method to assess the overall fairness of multiple LLMs across various labels. Through experiments with 16 LLMs, we uncover pervasive inconsistency, bias, and imbalanced inaccuracy across models, underscoring severe LLM judicial unfairness. Particularly, LLMs display notably more pronounced biases on demographic labels, with slightly less bias on substance labels compared to procedure ones. Interestingly, increased inconsistency correlates with reduced biases, but more accurate predictions exacerbate biases. While we find that adjusting the temperature parameter can influence LLM fairness, model size, release date, and country of origin do not exhibit significant effects on judicial fairness. Accordingly, we introduce a publicly available toolkit containing all datasets and code, designed to support future research in evaluating and improving LLM fairness.

cs.CL↗

arXiv2025

A universal scaling law for active diffusion in complex media

Using granular experiments and computer simulations, we investigate the long-time diffusion of active tracers in a broad class of complex media composed of frozen obstacles of diverse structures. By introducing a dimensionless persistence length $Q = v_d τ_r / d_t$, we propose a modified scaling relation that independently collapses experimental and simulation results across active and passive particles, diverse media, and distinct propulsion mechanisms. Our results reveal a universal active diffusion-structure relation that holds across both equilibrium and nonequilibrium regimes, providing a simple predictive framework for active diffusion in complex environments.

cond-mat.soft↗

arXiv2025

LACE: Controlled Image Prompting and Iterative Refinement with GenAI for Professional Visual Art Creators

We present LACE, a hybrid Human-AI co-creative system integrated into Adobe Photoshop supporting turn-taking and parallel interaction modes for iterative image generation. Through a study with 21 participants across representational, abstract, and design tasks, we found turn-taking preferred in early stages for idea generation, and parallel modes suited for detailed refinement. While this shorter workshop paper provides key insights and highlights, the comprehensive findings and detailed analysis are presented in a longer version available separately on arXiv.

cs.HC↗

arXiv2025

J&H: Evaluating the Robustness of Large Language Models Under Knowledge-Injection Attacks in Legal Domain

As the scale and capabilities of Large Language Models (LLMs) increase, their applications in knowledge-intensive fields such as legal domain have garnered widespread attention. However, it remains doubtful whether these LLMs make judgments based on domain knowledge for reasoning. If LLMs base their judgments solely on specific words or patterns, rather than on the underlying logic of the language, the ''LLM-as-judges'' paradigm poses substantial risks in the real-world applications. To address this question, we propose a method of legal knowledge injection attacks for robustness testing, thereby inferring whether LLMs have learned legal knowledge and reasoning logic. In this paper, we propose J&H: an evaluation framework for detecting the robustness of LLMs under knowledge injection attacks in the legal domain. The aim of the framework is to explore whether LLMs perform deductive reasoning when accomplishing legal tasks. To further this aim, we have attacked each part of the reasoning logic underlying these tasks (major premise, minor premise, and conclusion generation). We have collected mistakes that legal experts might make in judicial decisions in the real world, such as typos, legal synonyms, inaccurate external legal statutes retrieval. However, in real legal practice, legal experts tend to overlook these mistakes and make judgments based on logic. However, when faced with these errors, LLMs are likely to be misled by typographical errors and may not utilize logic in their judgments. We conducted knowledge injection attacks on existing general and domain-specific LLMs. Current LLMs are not robust against the attacks employed in our experiments. In addition we propose and compare several methods to enhance the knowledge robustness of LLMs.

cs.CL↗

arXiv2024

Glass Transition in Monolayers of Rough Colloidal Ellipsoids

Structure-dynamics correlation is one of the major ongoing debates in the glass transition, although a number of structural features have been found connected to the dynamic heterogeneity in different glass-forming colloidal systems. Here using colloidal experiments combined with coarse-grained molecular dynamics simulations, we investigate the glass transition in monolayers of rough colloidal ellipsoids. Compared with smooth colloidal ellipsoids, the surface roughness of ellipsoids is found to significantly change the nature of glass transition. In particular, we find that the surface roughness induced by coating only a few small hemispheres on the ellipsoids can eliminate the existence of orientational glass and the two-step glass transition found in monolayers of smooth ellipsoids. This is due to the surface roughness-induced coupling between the translational and rotational degrees of freedom in colloidal ellipsoids, which also destroys the structure-dynamics correlation found in glass-forming suspensions of colloidal ellipsoids. Our results not only suggest a new way of using surface roughness to manipulate the glass transition in colloidal systems, but also highlight the importance of detailed particle shape on the glass transition and structure-dynamics correlation in suspensions of anisotropic colloids.

cond-mat.soft↗

arXiv2024

Migration of active particle in mixtures of rigid and flexible rings

The migration of active particles in slowly moving, crowded, and heterogeneous media is fundamental to various biological processes and technological applications, such as cargo transport. In this study, we numerically investigate the dynamics of a single active particle in a medium composed of mixtures of rigid and flexible rings. We observe a non-monotonic dependence of diffusivity on the relative fraction of rigid to flexible rings, leading to the identification of an optimal composition for enhanced diffusion. This long-time non-monotonic diffusion, likely resulting from the different responses of the active particle to rigid and flexible rings, is coupled with transient short-time trapping. The probability distribution of trapping durations is well described by the extended entropic trap model. We further establish a universal relationship between particle activity and the optimal rigid-to-flexible ring ratio for diffusion, which aligns closely with our numerical results.

cond-mat.soft↗

arXiv2024

STARD: A Chinese Statute Retrieval Dataset with Real Queries Issued by Non-professionals

Statute retrieval aims to find relevant statutory articles for specific queries. This process is the basis of a wide range of legal applications such as legal advice, automated judicial decisions, legal document drafting, etc. Existing statute retrieval benchmarks focus on formal and professional queries from sources like bar exams and legal case documents, thereby neglecting non-professional queries from the general public, which often lack precise legal terminology and references. To address this gap, we introduce the STAtute Retrieval Dataset (STARD), a Chinese dataset comprising 1,543 query cases collected from real-world legal consultations and 55,348 candidate statutory articles. Unlike existing statute retrieval datasets, which primarily focus on professional legal queries, STARD captures the complexity and diversity of real queries from the general public. Through a comprehensive evaluation of various retrieval baselines, we reveal that existing retrieval approaches all fall short of these real queries issued by non-professional users. The best method only achieves a Recall@100 of 0.907, suggesting the necessity for further exploration and additional research in this area. All the codes and datasets are available at: https://github.com/oneal2000/STARD/tree/main

cs.IR↗

arXiv2023

On multi-step extended maximum residual Kaczmarz method for solving large inconsistent linear systems

A multi-step extended maximum residual Kaczmarz method is presented for the solution of the large inconsistent linear system of equations by using the multi-step iterations technique. Theoretical analysis proves the proposed method is convergent and gives an upper bound on its convergence rate. Numerical experiments show that the proposed method is effective and outperforms the existing extended Kaczmarz methods in terms of the number of iteration steps and the computational costs.

math.NA↗

arXiv2023

Random organization and non-equilibrium hyperuniform fluids on a sphere

Random organizing hyperuniform fluid induced by reciprocal activation is a non-equilibrium fluid with vanishing density fluctuations at large length scales like crystals. Here we extend this new state of matter to a closed manifold, namely a spherical surface. We find that the random organization on a spherical surface behaves similar to that in two dimensional Euclidean space, and the absorbing transition on a sphere also belongs to the conserved directed percolation universality class. Moreover, the reciprocal activation can also induce a non-equilibrium hyperuniform fluid on a sphere. The spherical structure factor at the absorbing transition and the non-equilibrium hyperuniform fluid phases are scaled as $S(l \rightarrow 0) \sim (l/R)^{0.45}$ and $S(l \rightarrow 0) \sim l(l+1)/R^2$, respectively, which are both hyperuniform according to the definition of hyperuniformity on a sphere with $l$ the wave number and $R$ the radius of the spherical surface. We also consider the impact of inertia in realistic hyperuniform fluids, and it is found only adding an extra length-scale, above which hyperuniform scaling appears. Our finding suggests a new method for creating non-equilibrium hyperuniform fluids on closed manifolds to avoid boundary effects.

cond-mat.soft↗

arXiv2022

On fast greedy block Kaczmarz methods for solving large consistent linear systems

A class of fast greedy block Kaczmarz methods combined with general greedy strategy and average technique are proposed for solving large consistent linear systems. Theoretical analysis of the convergence of the proposed method is given in detail. Numerical experiments show that the proposed methods are efficient and faster than the existing methods.

math.NA↗

arXiv2022

Restarted randomized surrounding methods for solving large linear equations

A class of restarted randomized surrounding methods are presented to accelerate the surrounding algorithms by restarted techniques for solving the linear equations. Theoretical analysis prove that the proposed method converges under the randomized row selection rule and the expectation convergence rate is also addressed. Numerical experiments further demonstrate that the proposed algorithms are efficient and outperform the existing method for over-determined and under-determined linear equation, as well as in the application of image processing.

math.NA↗

arXiv2021

Closing the B-tree vs. LSM-tree Write Amplification Gap on Modern Storage Hardware with Built-in Transparent Compression

This paper studies the design of B-tree that can take full advantage of modern storage hardware with built-in transparent compression. Recent years have witnessed significant interest in applying log-structured merge tree (LSM-tree) as an alternative to B-tree. The current consensus is that, compared with B-tree, LSM-tree has distinct advantages in terms of storage space efficiency and write amplification. This paper argues that one should revisit this belief upon the arrival of storage hardware with built-in transparent compression. Advanced storage appliances~(e.g., all-flash array) and emerging computational storage drives perform hardware-based lossless data compression, transparent to OS and user applications. Beyond straightforwardly reducing the physical storage cost difference between B-tree and LSM-tree, such modern storage hardware brings new opportunities to innovate B-tree implementation in order to largely reduce its write amplification. As the first step to explore the potential, this paper presents three simple design techniques (i.e., deterministic page shadowing, localized page modification logging, and sparse redo logging) that can leverage such modern storage hardware to significantly reduce the B-tree write amplification. We implemented these design techniques and carried out experiments on a commercial storage drive with built-in transparent compression. The results show that the proposed design techniques can reduce the B-tree write amplification by over 10x. Compared with RocksDB (a popular key-value store built upon LSM-tree), the implemented B-tree can achieve similar or even smaller write amplification and physical storage space usage.

cs.DB↗

arXiv2021

Kaczmarz-type inner-iteration preconditioned flexible GMRES methods for consistent linear systems

We propose using greedy and randomized Kaczmarz inner-iterations as preconditioners for the right-preconditioned flexible GMRES method to solve consistent linear systems, with a parameter tuning strategy for adjusting the number of inner iterations and the relaxation parameter. We also present theoretical justifications of the right-preconditioned flexible GMRES for solving consistent linear systems. Numerical experiments on overdetermined and underdetermined linear systems show that the proposed method is superior to the GMRES method preconditioned by NE-SOR inner iterations in terms of total CPU time.

math.NA↗

arXiv2020

Graph Regularized Nonnegative Tensor Ring Decomposition for Multiway Representation Learning

Tensor ring (TR) decomposition is a powerful tool for exploiting the low-rank nature of multiway data and has demonstrated great potential in a variety of important applications. In this paper, nonnegative tensor ring (NTR) decomposition and graph regularized NTR (GNTR) decomposition are proposed, where the former equips TR decomposition with local feature extraction by imposing nonnegativity on the core tensors and the latter is additionally able to capture manifold geometry information of tensor data, both significantly extend the applications of TR decomposition for nonnegative multiway representation learning. Accelerated proximal gradient based methods are derived for NTR and GNTR. The experimental result demonstrate that the proposed algorithms can extract parts-based basis with rich colors and rich lines from tensor objects that provide more interpretable and meaningful representation, and hence yield better performance than the state-of-the-art tensor based methods in clustering and classification tasks.

cs.CV↗