arXiv subjects

Kangkang Wang

Publications and source records attributed to Kangkang Wang.

6 recordsLinked to original sources

arXiv2026

MHPR: Multidimensional Human Perception and Reasoning Benchmark for Large Vision-Languate Models

Multidimensional human understanding is essential for real-world applications such as film analysis and virtual digital humans, yet current LVLM benchmarks largely focus on single-task settings and lack fine-grained, human-centric evaluation. In this work, we introduce MHPR, a comprehensive benchmark for joint perception-reasoning over human-centric scenes spanning individual, multi-person, and human-object interaction dimensions. MHPR comprises a multi-level data design-Captioned Raw Data (C-RD), Supervised Fine-Tuning Data (SFT-D), Reinforcement Learning Data (RL-D), and Test Data (T-D)-together with an automated caption/VQA generation pipeline (ACVG) that performs category-wise attribute decomposition, attribute-specific rewriting, and multi-model voting to ensure high-quality, scalable annotations. We evaluate state-of-the-art vision-language models on fine-grained attributes (appearance, clothing, pose, parts) and high-level semantics (social relations, action semantics, spatial relations, intent and functionality). Our findings show that: 1) format-aligned SFT data substantially improves instruction following and stability; 2) challenge-focused RL data derived from bad-case analysis further enhances perception and reasoning on difficult instances; and 3) training Qwen2.5-VL-7B with MHPR yields significant gains, achieving near-parity with considerably larger models. We release ACVG and MHPR to facilitate reproducible, extensible research on human-centric perception and reasoning.

cs.CV↗

arXiv2025

Supersonic wave propagation in active non-Hermitian acoustic metamaterials

Obtaining a group velocity higher than the speed of sound in a waveguide is a challenging task in acoustic wave engineering. Even more challenging is to achieve this velocity increase without any intervention with the waveguide profile, such as narrowing or widening, and particularly without interfering with the passage by flexible inclusions, either passive or active. Here, we approach this problem by invoking concepts from non- Hermitian physics, and imposing them using active elements that are smoothly sealed within the waveguide wall. In a real-time feedback operation, the elements induce local pressure gain and loss, as well as non-local pressure integration couplings. We employ a dedicated balancing between the control parameters, derived from lattice theory and adjusted to the waveguide system, to drive the dynamics into a stable parity-time-symmetric regime. We demonstrate the accelerated propagation of a wave packet both numerically and experimentally in an air-filled waveguide and discuss the trade-off between stabilization and the achievable velocity increase. Our work prepares the grounds for advanced forms of wave transmission in continuous media, enabled by short and long range active couplings, created via embedded real-time feedback control.

physics.app-ph↗

arXiv2021

Semantically Robust Unpaired Image Translation for Data with Unmatched Semantics Statistics

Many applications of unpaired image-to-image translation require the input contents to be preserved semantically during translations. Unaware of the inherently unmatched semantics distributions between source and target domains, existing distribution matching methods (i.e., GAN-based) can give undesired solutions. In particular, although producing visually reasonable outputs, the learned models usually flip the semantics of the inputs. To tackle this without using extra supervision, we propose to enforce the translated outputs to be semantically invariant w.r.t. small perceptual variations of the inputs, a property we call "semantic robustness". By optimizing a robustness loss w.r.t. multi-scale feature space perturbations of the inputs, our method effectively reduces semantics flipping and produces translations that outperform existing methods both quantitatively and qualitatively.

cs.CV↗

arXiv2019

Federated Evaluation of On-device Personalization

Federated learning is a distributed, on-device computation framework that enables training global models without exporting sensitive user data to servers. In this work, we describe methods to extend the federation framework to evaluate strategies for personalization of global models. We present tools to analyze the effects of personalization and evaluate conditions under which personalization yields desirable models. We report on our experiments personalizing a language model for a virtual keyboard for smartphones with a population of tens of millions of users. We show that a significant fraction of users benefit from personalization.

cs.LG↗

arXiv2017

Mirror Symmetry and Discriminants

We analyze the locus, together with multiplicities, of "bad" conformal field theories in the compactified moduli space of N=(2,2) superconformal field theories in the context of the generalization of the Batyrev mirror construction using the gauged linear sigma-model. We find this discriminant of singular theories is described beautifully by the GKZ "A-determinant" but only if we use a noncompact toric Calabi-Yau variety on the A-model side and logarithmic coordinates on the B-model side. The two are related by "local" mirror symmetry. The corresponding statement for the compact case requires changing multiplicities in the GKZ determinant. We then describe a natural structure for monodromies around components of this discriminant in terms of spherical functors. This can be considered a categorification of the GKZ A-determinant. Each component of the discriminant is naturally associated with a category of massless D-branes.

hep-th↗

arXiv2016

First order reversal curves and intrinsic parameter determination for magnetic materials; limitations of hysteron-based approaches in correlated systems

The generic problem of extracting information on intrinsic particle properties from the whole class of interacting magnetic fine particle systems is a long standing and difficult inverse problem. As an example, the Switching Field Distribution (SFD) is an important quantity in the characterization of magnetic systems, and its determination in many technological applications, such as recording media, is especially challenging. Techniques such as the first order reversal curve (FORC) methods, were developed to extract the SFD from macroscopic measurements. However, all methods rely on separating the contributions to the measurements of the intrinsic SFD and the extrinsic effects of magnetostatic and exchange interactions. We investigate the underlying physics of the FORC method by applying it to the output predictions of a kinetic Monte-Carlo model with known input parameters. We show that the FORC method is valid only in cases of weak spatial correlation of the magnetization and suggest a more general approach.

cond-mat.mtrl-sci↗