arXiv:2605.15638v1 Announce Type: new
Abstract: Hyperscaler reports of silent data corruptions (SDCs), presumed to be caused by silicon manufacturing defects, have motivated the development of functional tests for detecting defective CPUs. We present ITHICA, an approach for automatically generating functional tests for defect-induced errors from arbitrary programs by inserting intra-thread, instruction-level error checks, primarily leveraging instruction duplication and output comparison. Our key insight is that the most pernicious defects cause inconsistent errors: two executions of the same instruction within the same thread, given the same inputs, can produce different architectural outputs depending on the execution context in which they run. By exploiting this insight, ITHICA enables arbitrary programs to serve as tests and identifies affected instructions upon error detections. We use ITHICA to transform industrial hyperscaler test programs (our baseline), datacenter workloads, and common libraries into functional tests, and evaluate them on over 3,000 CPU servers. ITHICA error checks detect 39% more defective servers than native checks within the ITHICA tests derived from our baseline programs, and enable novel findings on defect behavior that challenge conclusions drawn by prior hyperscaler fleet studies.
Science Journals
arXiv:2601.21294v2 Announce Type: replace
Abstract: Partial Least Squares (PLS) learns shared structure from paired data via the top singular vectors of the empirical cross-covariance (PLS-SVD), but multimodal datasets often have missing entries in both views. We study PLS-SVD under independent entry-wise missing-completely-at-random masking in a proportional high-dimensional spiked model. After appropriate normalization, the masked cross-covariance behaves like a spiked rectangular random matrix whose effective signal strength is attenuated by $\sqrt{\rho}$, where $\rho$ is the joint entry retention probability. The replica-symmetric analysis predicts a sharp BBP-type phase transition: below a critical signal-to-noise threshold the leading singular vectors are asymptotically uninformative, while above it they achieve nontrivial alignment with the latent shared directions, with closed-form asymptotic overlap formulas. We also state a finite-rank extension as a conjecture, predicting that the same missingness-adjusted threshold applies componentwise when the latent spikes are separated. Simulations and semi-synthetic multimodal experiments agree with the predicted phase diagram and recovery curves across aspect ratios, signal strengths, and missingness levels.
arXiv:2605.15637v1 Announce Type: new
Abstract: Motivated by the prospect of chiral-mode control in compact photonic systems, we analyze discrete coupled single-mode resonators. Using the minimal three-resonator model, we show that an infinitesimal complex onsite perturbation near a Hermitian diabolic point (DP) induces chiral-mode selection, governed by what we term an asymptotic exceptional point (AEP). Here, an AEP denotes a Hermitian DP equipped with a non-Hermitian perturbation that induces an asymptotically defective effective Hamiltonian. The eigenvectors coalesce in the asymptotic limit toward the DP, although the Hamiltonian at the point itself remains diagonalizable. Operationally, this AEP response realizes chirality switching from an achiral state to a chiral state. The associated eigenvalue response exhibits the anomalous fractional-power scaling ${\Delta}{\lambda} \propto {\varepsilon}^{3/2}$, distinct from the square-root response of an ordinary exceptional point (EP). We further show that, in a broader two-parameter perturbation space, ordinary EPs lie on exceptional-line branches that meet at the AEP. A finitebias control sweep crosses these branches at an EP pair, enabling chirality reversal between opposite chiral states. The central message is therefore that the AEP organizes two related routes for chirality switching: direct switching from an achiral state to a chiral state via the AEP, and switching between opposite chiral states via an EP pair in the vicinity of the AEP. Within a finite-resolution averaging model, these two operating points exhibit different practical performance characteristics, and under sufficiently high control resolution, the AEP operating point can become more favorable than the EP-pair operating point, suggesting a route toward compact and low-energy chiral photonic devices.
arXiv:2605.15890v1 Announce Type: new
Abstract: We propose a communication-efficient optimally structured gradient coding scheme to jointly address straggler resilience and communication efficiency in heterogeneous distributed learning. By establishing a unified framework that simultaneously optimizes gradient coding and quantization, we formulate an optimization problem to minimize residual error subject to an unbiasedness constraint. We rigorously establish the joint global optimum by deriving a closed-form code structure coupled with an optimal bit allocation strategy, while simultaneously proposing a low-complexity bit allocation algorithm that efficiently yields near-optimal performance. We provide rigorous convergence analysis for convex and smooth functions. Experiments on the COCO dataset demonstrate that our joint design significantly accelerates convergence and enhances communication efficiency compared to existing baselines.
arXiv:2605.15714v1 Announce Type: new
Abstract: This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established "shift-left" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.
arXiv:2605.15923v1 Announce Type: new
Abstract: Modern image encoders achieve high generalization by decoupling semantic meaning from resolution, an ability yet to be fully realized in the 3D domain. We investigate the failure of 3D point cloud encoders to achieve similar generalization and find that existing models are highly sensitive to sampling resolution and scale changes, leading to significant performance degradation. This sensitivity is a major bottleneck for real-world deployment in robotics, as it suggests models overfit to specific quantization densities and object scales rather than learning invariant semantic features. To mitigate this dependency, we propose Invaria, a point cloud encoder that achieves scale and density invariance through next-resolution prediction and receptive field calibration. While our objective is not the explicit generation of high-resolution point clouds, we find that this training objective encourages the model to learn robust, structural invariants. The resulting encoder achieves significant performance gains during resolution shifts while maintaining high efficiency through a compact model size and reduced token requirements. Specifically, on ScanNet, Invaria achieves a 56.0\% higher mIoU at 3$\times$ lower resolution and a 20\% improvement when the objects scale is reduced by a factor of 3. These gains are achieved with a 45\% smaller model size and an average reduction of 40\% in input tokens.
arXiv:2605.15636v1 Announce Type: new
Abstract: This note deals with a tearing and interconnecting (special non-overlapping domain decomposition) formulation for magneto-quasi-statics (also known as the eddy current model). Only two subdomains are considered, one conducting and one insulating. Using a straightforward tree-cotree splitting, one can get rid of some kernel components in the non-conducting region, but due to the coupling across the interface, a lot of kernel functions remain that are associated with the interface. The formulation presented here overcomes this problem by using a space splitting into gradient fields and a complementary space. Under a mild condition on that splitting, it is shown that (i) one does not need any gradient part in the non-conducting domain, and therefore no coupling of any gradient components between the two subdomains, (ii) both subdomain operators are invertible, and (iii) although the magnetic vector potential is discontinuous across the subdomain interface, the corresponding magnetic field is globally in H(div).
arXiv:2605.15382v1 Announce Type: new
Abstract: Low-rank methods for kinetic equations have attracted increasing attention due to their effectiveness in reducing the high dimensionality of phase space. In our previous work [G. Wang & J. Hu, J. Comput. Phys. 558 (2026) 114884], we developed a dynamical low-rank method based on the projector-splitting integrator in tensor-train (TT) format, in which explicit time integration is employed in all substeps. As a result, the method is subject to severe stability constraints in the strongly collisional regimes. In this paper, we consider kinetic equations with the (nonlinear) Fokker--Planck collision operator and develop a dynamical low-rank method that employs implicit or implicit-explicit (IMEX) discretizations in appropriate substeps to overcome stiffness. In these implicit substeps, the resulting equations can be formulated as matrix or tensor Sylvester equations, for which we propose efficient direct solvers by exploiting their underlying structure. The overall computational cost of the proposed method scales linearly with respect to the number of grid points in a single velocity dimension, comparable to that of a fully explicit low-rank scheme. We demonstrate the accuracy and efficiency of the proposed method on several representative kinetic test problems.
arXiv:2605.15924v1 Announce Type: new
Abstract: In virtual reality environments, the alignment of perceptual modalities is crucial for immersion and presence. In the AR domain, it is difficult to create such alignments because elements in the physical world are often beyond the user's control. However, recent advances in generative AI enable on-demand content creation, enabling highly reactive AR experiences. Combined with contextual information about the physical world, it has become possible to design experiences that seamlessly align with the user's environment. In this reflection paper, I emphasize the importance of "synchronized" realities for context-aware AR experiences, particularly in mobility scenarios. I present several examples of existing synchronized experiences and examine their commonalities and distinctions. Finally, I discuss opportunities and pitfalls of synchronizing AR experiences with the physical world.
arXiv:2605.15925v1 Announce Type: new
Abstract: Let $\mathbb{F}_{p^m}$ be the field containing $p^m$ elements where $p$ is an odd prime and $m \in \mathbb{N}$.
In this article, we propose a unified approach to the study of skew constacyclic codes of length $np^s$ over the ring $R_k = \mathbb{F}_{p^m}[u]/\langle u^k \rangle,$ where $n, s, k \in \mathbb{N}$ and $\gcd(n, p)=1$.
Consider the skew polynomial ring $R_k[x;\Theta]$, where
$\Theta$ is an automorphism of $R_k$ such that $xa = \Theta(a)x$ for all $a \in R_k$. Let $f(x)$ be a central irreducible divisor of $x^{np^s} - \lambda$ of degree $l$ and multiplicity $j$ in $R_k[x;\Theta]$, where $\lambda $ is an invertible element in $R_k$. In this article, we study skew constacyclic codes of length \(np^s\) over \(R_k\), which reduces to the study of skew polycyclic codes of length $jl$ associated with a polynomial \(f(x)^j\).
Using the fact that skew polycyclic codes associated with a polynomial \(f(x)^j\) can be described by the left ideal structure of the quotient ring $R_k[x;\Theta]/\langle f(x)^{j}\rangle$, we investigate this class of codes for specific choices of $\Theta$. In particular, if $\lambda$ is an invertible element of $\mathbb{F}_{p^m}$, we classify all left ideals and establish an isomorphism between skew cyclic and skew constacyclic codes, under suitable conditions.
Furthermore, we provide a comprehensive analysis of skew constacyclic codes of length $3p^s$ over $R_k$. Finally, we examine skew cyclic and skew negacyclic codes of length $6p^s$ over $R_k$ using the factorization of $x^{6p^s} - 1$ and $x^{6p^s} + 1$, respectively; with a complete case-by-case analysis. Examples demonstrating codes with optimal parameters are also included.
arXiv:2508.15601v2 Announce Type: replace
Abstract: Mixed-precision inference techniques reduce the memory and computational demands of Large Language Models (LLMs) by applying hybrid precision formats to model weights, activations, and KV caches. However, existing systems struggle to (i) automatically generalize across diverse hardware architectures and precision formats, often requiring fragmented, hand-tuned kernels, and (ii) fully exploit available memory and compute resources, often causing performance bottlenecks. To address these problems, we propose TurboMind, a generalizable and efficient mixed-precision LLM inference engine of LMDeploy. TurboMind is built around two hardware-aware mixed-precision pipelines: A General Matrix Multiply (GEMM) pipeline that optimizes matrix operations through offline weight packing and online acceleration, and an attention pipeline that enables efficient attention computation with different Query, Key, and Value precision combinations. These pipelines are enabled by four key techniques: (i) Hardware-aware weight packing and (ii) adaptive head alignment for generalizability, and (iii) instruction-level parallelism and (iv) a KV memory loading pipeline for efficiency. We conduct comprehensive evaluations of LMDeploy powered by TurboMind across sixteen popular LLMs and four representative GPU architectures. Results demonstrate that LMDeploy achieves up to 61% lower serving latency (30% on average) and up to 156% higher throughput (58% on average) in mixed-precision workloads compared to existing mixed-precision frameworks, establishing consistent performance improvements across all tested configurations and hardware types. This work is open-sourced and publicly available at https://github.com/InternLM/lmdeploy.
arXiv:2605.15635v1 Announce Type: new
Abstract: Linguistic ambiguity is critical to the robustness of Large Language Models (LLMs), yet existing research focuses mostly on English, with limited attention devoted to Chinese. Existing Chinese ambiguity datasets (e.g., CHAmbi) suffer from poor scalability. Guided by Potential Ambiguity (PA) Theory, we design a semi-automatic pipeline to construct CHA-Gen. It is the first PA Theory-grounded Chinese ambiguity dataset, which comprises 5,712 sentences (2,414 ambiguous, 3,298 unambiguous) across 18 potential ambiguous structures. Evaluating LLMs (e.g. Gemma 3, Qwen 2.5/3 series) via direct querying and machine translation, we find that LLMs struggle with ambiguity detection (improved by CoT prompting). Analysis of Qwen3-32B's CoT rationales reveals three common failure modes: ambiguity blindness, misattribution, and premature resolution. Uncertainty quantification with semantic entropy metric shows higher uncertainty for ambiguous sentences. Moreover, instruction tuning induces overconfidence, whereas Base models better capture semantic diversity. We further observe that models exhibit a bias toward dominant interpretations. Our work provides a scalable approach for Chinese ambiguity corpus and insights into LLMs' ambiguity handling, laying a foundation for enhancing Chinese ambiguity research in LLMs.
arXiv:2605.15632v1 Announce Type: new
Abstract: Time-reflection occurs when a wave is propagating in a medium undergoing a large and abrupt change in its properties: the original wave splits into a time-refracted wave and a time-reflected wave, each displaying different features. The time-refracted wave continues along its original course but experiences a frequency shift, whereas the time-reflected wave is propagating backwards in space with a reversed phase, also with a shifted frequency. These phenomena are fundamental to any wave system, but the most interesting are electromagnetic (EM) waves, specifically at optical frequencies, where they can couple to light-matter interactions. However, time-reflection of EM waves was thus far observed only at RF frequencies, never at optical frequencies. This is because time-reflection requires an order-unity variation of the refractive index occurring faster than a single wave cycle, and conventional optical nonlinearities are either too weak or too slow by orders of magnitude. Here, we present the first observation of time-reflection at optical frequencies. We induce an order-unity refractive-index change with sub-cycle duration, observe the time-reflection, and study its fundamental properties. These results provide an experimental pathway to experimenting with time-interfaces, generating photonic time-crystals and exploring new regimes of light-matter interaction in time-varying media.
arXiv:2604.08302v3 Announce Type: replace
Abstract: We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMax
arXiv:2605.15713v1 Announce Type: new
Abstract: Legged manipulators extend robotic capabilities beyond static manipulation by integrating agile locomotion with versatile arm control. However, achieving precise manipulation while maintaining coordinated locomotion remains a major challenge. This work presents a hierarchical reinforcement learning framework for dynamic pick-and-place tasks using a quadruped equipped with a 6-DOF robotic arm. The framework incorporates an explicit mass estimation module enabling adaptive whole-body control for objects with varying weights. In simulation, the system achieves an 86.05% success rate with payloads up to 2.3 kg. The approach is further validated through real-world experiments across six representative scenarios with controlled variations in object physical properties (size and mass) and task heights. Specifically, within a wide vertical workspace ranging from ground level to 1.1~m-high tabletops, the system demonstrates an average success rate of 73.3% for payloads up to 1.3 kg, with an average execution time of 4.06 s. Unlike prior works that handle lightweight objects and execute pick-and-place motions with slow, piecewise motions, the proposed framework exploits concurrent locomotion and manipulation for dynamic, continuous execution. These results demonstrate the potential of quadrupedal mobile manipulators for adaptive, whole-body pick-and-place with heavier payloads and extended workspaces.
arXiv:2605.08642v2 Announce Type: replace
Abstract: We develop a descriptive account of scientific reward in physics based on the concept of the time-dependent Polydoxon, defined as the structured set of empirically viable theories at a given time. We argue that highly rewarded contributions, such as those recognized by major prizes and professional honors, can be systematically understood as those that transform this space. These transformations take the form of expansion (adding viable theories), contraction (eliminating viable theories), reconfiguration (illuminating deeper structures and relations within and between theories), and enabling moves (methodological or technological advances that enable future transformations). The analysis is further refined by emphasizing that reward correlates with the transformation's magnitude, assessed along dimensions of scope, centrality, depth, and future leverage. This framework reframes the analysis of rewarded achievement away from isolated theoretical successes and toward the dynamics of a landscape of viable theories, providing a more unified descriptive interpretation of rewarded scientific activity in physics across its diverse set of theoretical and experimental discoveries.
arXiv:2605.15933v1 Announce Type: new
Abstract: We first show how the cohomology of some Bernstein-Gelfand-Gelfand (BGG) sequences that are important for the numerical analysis of partial differential equations, can be obtained through the construction of a long exact sequence connecting cohomology groups. Then we explain the extension of this result to the non-injective/surjective case through the systematic use of short exact sequences of complexes and their associated long exact sequences of cohomology groups. Finally an interpretation in terms of spectral sequences is given.
arXiv:2605.15549v1 Announce Type: new
Abstract: The demand for clean energy is ever increasing, with new nuclear technologies presenting a complementary solution to renewable energies. However, designing and operating these systems is exceptionally difficult, given the complexity of the physical phenomena that interact to form the system dynamics. While high-fidelity simulations help to understand the non-linear, multi-physics interactions within a reactor, they are computationally expensive and rarely suitable for real-time applications. Furthermore, model-based approaches are inherently sensitive to simplifying assumptions required to derive their governing equations and parameters, leading to inevitable discrepancies with real-world measurements. In contrast, Machine Learning (ML) methods have the potential to generate reliable surrogate models which may be able to quickly predict the system's behaviour. However, the number of data-driven methods that can potentially be used for this task is large and diverse. In a safety-critical setting such as nuclear engineering, a fair comparison of different ML methods, and a clear understanding of their advantages and limitations, is of paramount importance. To address this, we introduce a Common Task Framework (CTF) for ML in nuclear engineering, building upon previous efforts in dynamical systems and seismology. This CTF considers a curated set of datasets from different nuclear and nuclear-adjacent systems. The CTF evaluates the performance of a method on 12 established metrics, alongside a new paradigm focused on system monitoring from sparse measurements only. We illustrate the framework by benchmarking standard ML baselines against these datasets, revealing current method limitations. Our vision is to replace ad hoc comparisons with standardized evaluations on hidden test sets, raising the bar for rigour and reproducibility in scientific ML for the nuclear industry.
arXiv:2605.15380v1 Announce Type: new
Abstract: Recent advances in generative AI have shown their potential to be leveraged for legal education. Yet, work on the development and deployment of such systems for legal education in the Global South is limited. In this work, we developed Eskwai for Students, a generative AI assistant to help law students with their legal education. Eskwai for Students is a retrieval augmented generation (RAG) system that provides answers to a wide range of legal questions for law students grounded in a curated database of over 12K case laws and 1.4K legislation in Ghana. We deployed Eskwai for Students in a longitudinal study of 30 months (2.5 years) used by 3.1K law students in Ghana who made 32K queries. We evaluated the helpfulness of our AI, and provided insight into the kinds of queries law students submit to this generative AI tool, which raises some ethical concerns. This work contributes to an understanding of how law students in the Global South are using generative AI for their studies and the ways it could be leveraged responsibly to advance legal education.
arXiv:2605.15630v1 Announce Type: new
Abstract: Free energy profiles serve as a fundamental bridge between microscopic atomic fluctuations and macroscopic thermodynamic observables. Estimating the free energy profile along a reaction coordinate, referred to as the potential of mean force (PMF), with density functional theory (DFT) accuracy is computationally expensive. Universal machine learning interatomic potentials (MLIPs) drastically reduce this cost, but their accuracy is strongly determined by their training data and hence can be uncertain for a given system. In this work, we present a systematic and scalable framework for reweighting PMFs, initially sampled with a single 'source' MLIP, across a representative suite of target MLIPs. Because traditional direct exponential reweighting fails for large system sizes due to low phase-space overlap between potentials, we deploy robust analytical corrections. Applying this to a complex 601-atom system of Li$^+$ transport in a nanoconfined electrolyte, we demonstrate that a mean energy-gap approximation effectively bypasses statistical collapse, producing a highly stable PMF matching the target PMF. Using this approach, we recover high-fidelity target thermodynamics across multiple DFT reference levels (PBE+D3, PBE-sol, r$^2$SCAN,r$^2$SCAN-D4) at a fraction of the computational cost of full simulations. Furthermore, thermodynamic analysis reveals that the studied MLIPs partition into two distinct clusters driven by their training data. Our reweighting framework successfully recovers target thermodynamic properties--specifically, reaction and activation free energies--even when the phase-space overlap between potentials is critically low. Ultimately, this approach establishes a vital diagnostic protocol to achieve affordable cross-model consensus on materials chemistry properties without redundant, resource-intensive simulations.
arXiv:2605.15627v1 Announce Type: new
Abstract: In this paper, we present an improved numerical algorithm for computing the intersection area of multiple circles and a complex polygon efficiently. This geometric problem is fundamental to applications such as wireless sensor networks and base station deployment. The key idea is a curvature-multiplicity-guided adaptive sampling strategy that dynamically concentrates sampling points in geometrically complex boundary regions. The algorithm integrates three components: (i) adaptive quadtree partitioning, (ii) analytical integration via Green's theorem for cells intersecting a single circle, and (iii) curvature-multiplicity-guided Monte Carlo subsampling for cells intersecting multiple circles, where a minimum sample count and a constant factor are introduced into the sampling size. Theoretical analysis shows that the algorithm achieves O(1/{\epsilon}3/2) computational complexity while maintaining an O({\epsilon}) error bound, improving upon the O(1/{\epsilon}2) complexity of classical Monte Carlo and uniform grid methods for the same error tolerance {\epsilon}. Numerical experiments on complex polygons, including synthetic data and real-world scenarios, demonstrate that our algorithm outperforms five classical methods in terms of relative error. Furthermore, parameter sensitivity analysis confirms that the algorithm is robust and could make it suited for practical applications such as wireless sensor network coverage estimation.
arXiv:2605.15484v1 Announce Type: new
Abstract: Mixture-of-Experts (MoE) networks promise favorable accuracy-compute trade-offs, yet practical vision deployments are hindered by expert collapse and limited end-to-end efficiency gains. We study when sparse top-$k$ routing with hard capacity constraints helps in vision classification, evaluated under multi-seed protocols on four benchmarks (CIFAR-10/100, Tiny-ImageNet, ImageNet-1K). We observe a \emph{compute-leverage pattern}: positive accuracy gaps require a substantial fraction $\rho$ of total FLOPs to be routed; at ImageNet scale this is necessary but not sufficient, as multi-expert routing ($k \geq 2$) is additionally required. Two controlled experiments isolate these factors. A hidden-size sweep on CIFAR-10 yields both predicted sign reversals across standard and depthwise backbones, ruling out backbone family as the active variable. An ImageNet-1K ablation that varies only top-$k$ -- holding architecture, initialization, and $\rho$ fixed -- reverses the gap from positive to negative across all five seeds. A per-sample variant of Soft MoE that softmaxes over experts rather than the batch rescues CIFAR-100 above the dense baseline, identifying batch-axis dispatch as the dominant failure mode in per-sample CNN settings. Code and aggregate results: https://github.com/libophd/sparse-moe-vision-rho.
arXiv:2605.15379v1 Announce Type: new
Abstract: The log-homotopy particle flow filter resolves the Bayesian update by transporting particles along a continuous trajectory in pseudo-time. However, the governing partial differential equation for the flow velocity is fundamentally underdetermined, admitting an infinite family of valid solutions. In this work, we regard the particle flow as the motion of a pressureless inviscid fluid. We define a Lagrangian action based on the kinetic energy of the system, subject to the constraints imposed by the continuity equation and the log-homotopy evolution. By applying the principle of least action, we obtain the Euler--Lagrange equations for the optimal flow, which yields an irrotational potential flow structure. We show that this variational framework yields a coupled Hamilton--Jacobi equation structurally isomorphic to Madelung's hydrodynamic formulation of quantum mechanics. In this analogy, the log-homotopy constraint acts as a generalized quantum potential that generates the force required to guide the probability fluid along the exact Bayesian update path. Finally, we derive the material acceleration of the flow, shifting the formulation from a kinematic to a dynamical description. This perspective could enable the application of higher-order symplectic integrators for improved numerical stability and provide a physics-based metric for adaptive stiffness detection in high-dimensional filtering.
arXiv:2605.15626v1 Announce Type: new
Abstract: Large language models deliver strong performance across language and reasoning tasks, but their storage and compute costs remain major barriers to deployment in resource-constrained and latency-sensitive settings. SVD-based post-training compression offers a hardware-agnostic way to reduce model size and improve inference efficiency through low-rank factorization. However, existing methods often rely on input-only whitening spaces, homogeneous rank allocation, or loss-agnostic allocation heuristics, limiting their ability to preserve model quality under aggressive compression. We propose Input-Output Whitened SVD (IO-SVD), a post-training compression method that forms a KL-aware double-sided whitening space for model weights. Using a second-order expansion of the KL loss over the top-K token probabilities, IO-SVD constructs an output-side metric that captures predictive sensitivity, while input whitening captures activation statistics. We further introduce an efficient heterogeneous rank-allocation strategy that scores whitened singular components using first-order calibration loss estimates and prunes the least sensitive components under a global budget. Inspired by prior work that combines SVD truncation with quantization, we improve hybrid SVD-quantization compression through loss-aware remapping, which selects low-rank factor rows for 8-bit quantization based on the predicted loss change incurred by quantizing them. Extensive experiments across diverse LLM and VLM families, and inference-time analysis shows that IO-SVD compresses LLMs with minimal performance degradation while delivering practical inference speedups. Code is available at https://github.com/mint-vu/IO-SVD.git
arXiv:2511.13108v4 Announce Type: replace
Abstract: The rapid progress of generative models such as GANs and diffusion models has led to the widespread proliferation of AI-generated images, raising concerns about misinformation, privacy violations, and trust erosion in digital media. Although large-scale multimodal models like CLIP offer strong transferable representations for detecting synthetic content, fine-tuning them often induces catastrophic forgetting, which degrades pre-trained priors and limits cross-domain generalization. To address this issue, we propose the Distillation-guided Gradient Surgery Network (DGS-Net), a novel framework that preserves transferable pre-trained priors while suppressing task-irrelevant components. Specifically, we introduce a gradient-space decomposition that separates harmful and beneficial descent directions during optimization. By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net achieves unified optimization of prior preservation and irrelevant suppression. Extensive experiments on 50 generative models demonstrate that our method outperforms state-of-the-art approaches by an average margin of 6.6, achieving superior detection performance and generalization across diverse generation techniques.