Searcharxiv⌕ Search

arXiv subjects

Yi-Ting Li

Publications and source records attributed to Yi-Ting Li.

5 recordsLinked to original sources

Gradient-Guided Hardware-Native MIXFP4

Blackwell tensor cores support native inference with 4-bit weights and activations (W4A4), but NVFP4 leaves a perplexity (PPL) gap to BF16. Choosing a format for each 16-element block can improve accuracy, but current matrix multiply instructions cannot execute choices this fine. Applying these rules at hardware-compatible tiles loses most of their benefit. We introduce FlipQuant, which selects E2M1 or E0M3 for each MMA weight tile. It keeps both quantized weight candidates fixed and learns a binary format map from end-to-end gradients of a bucketed KL loss to a BF16 teacher, re-evaluating them after each update. Native kernels execute 16 x 64 tiles on SM120 and 256 x 64 tiles on SM100 with median GEMM overheads of 1.9% and 1.7%, respectively. Across six language models and two corpora, FlipQuant at 16 x 64 closes 42.5% of NVFP4's PPL gap to BF16 on average, compared with 13.3% for FourOverSix. Code is available at https://github.com/brian030128/flipquant.

cs.LG↗

CFEVER: A Chinese Fact Extraction and VERification Dataset

We present CFEVER, a Chinese dataset designed for Fact Extraction and VERification. CFEVER comprises 30,012 manually created claims based on content in Chinese Wikipedia. Each claim in CFEVER is labeled as "Supports", "Refutes", or "Not Enough Info" to depict its degree of factualness. Similar to the FEVER dataset, claims in the "Supports" and "Refutes" categories are also annotated with corresponding evidence sentences sourced from single or multiple pages in Chinese Wikipedia. Our labeled dataset holds a Fleiss' kappa value of 0.7934 for five-way inter-annotator agreement. In addition, through the experiments with the state-of-the-art approaches developed on the FEVER dataset and a simple baseline for CFEVER, we demonstrate that our dataset is a new rigorous benchmark for factual extraction and verification, which can be further used for developing automated systems to alleviate human fact-checking efforts. CFEVER is available at https://ikmlab.github.io/CFEVER.

cs.CL↗

Homoclinic orbits in Kerr-Newman black holes

We present the exact solutions of the homoclinic orbits for the timelike geodesics of the particle on the general nonequatorial orbits in the Kerr-Newman black holes. The homoclinic orbit is the separatrix between bound and plunging geodesics, a solution that asymptotes to an energetically bound, unstable spherical orbit. The solutions are written in terms of the elliptical integrals and the Jacobi elliptic functions of manifestly real functions of the Mino time where we focus on the effect from the charge of the black hole to the homoclinic orbits. The parameter space of the homoclinic solutions is explored. The nonequatorial homoclinic orbits in Kerr cases can be obtained by setting the charge of the black holes to be zero. The homoclinic orbits and the associated phase portrait as a function of the radial position and its derivation with respect to the Mino time are plotted using the analytical solutions. In particular, the solutions can reduce to the zero azimuthal angular moment homoclinic orbits for understanding the frame dragging effects from the spin as well as the charge of the black hole. The implications of the obtained results to observations are discussed.

gr-qc↗

A note on eigenvalues of random block Toeplitz matrices with slowly growing bandwidth

This paper can be thought of as a remark of \cite{llw}, where the authors studied the eigenvalue distribution $μ_{X_N}$ of random block Toeplitz band matrices with given block order $m$. In this note we will give explicit density functions of $\lim\limits_{N\to\infty}μ_{X_N}$ when the bandwidth grows slowly. In fact, these densities are exactly the normalized one-point correlation functions of $m\times m$ Gaussian unitary ensemble (GUE for short). The series $\{\lim\limits_{N\to\infty}μ_{X_N}|m\in\mathbb{N}\}$ can be seen as a transition from the standard normal distribution to semicircle distribution. We also show a similar relationship between GOE and block Toeplitz band matrices with symmetric blocks.

math.PR↗

Limit Distributions of Eigenvalues for Random Block Toeplitz and Hankel Matrices

Block Toeplitz and Hankel matrices arise in many aspects of applications. In this paper, we will research the distributions of eigenvalues for some models and get the semicircle law. Firstly we will give trace formulae of block Toeplitz and Hankel matrix. Then we will prove that the almost sure limit $γ_{_T}^{(m)}$ $(γ_{_H}^{(m)})$ of eigenvalue distributions of random block Toeplitz (Hankel) matrices exist and give the moments of the limit distributions where $m$ is the order of the blocks. Then we will prove the existence of almost sure limit of eigenvalue distributions of random block Toeplitz and Hankel band matrices and give the moments of the limit distributions. Finally we will prove that $γ_{_T}^{(m)}$ $(γ_{_H}^{(m)})$ converges weakly to the semicircle law as $m\ra\iy$.

math.PR↗