arXiv subjects

Jiaqi Liao

Publications and source records attributed to Jiaqi Liao.

25 records · Page 2Linked to original sources

arXiv2024

Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation

Text-to-video (T2V) models like Sora have made significant strides in visualizing complex prompts, which is increasingly viewed as a promising path towards constructing the universal world simulator. Cognitive psychologists believe that the foundation for achieving this goal is the ability to understand intuitive physics. However, the capacity of these models to accurately represent intuitive physics remains largely unexplored. To bridge this gap, we introduce PhyGenBench, a comprehensive \textbf{Phy}sics \textbf{Gen}eration \textbf{Ben}chmark designed to evaluate physical commonsense correctness in T2V generation. PhyGenBench comprises 160 carefully crafted prompts across 27 distinct physical laws, spanning four fundamental domains, which could comprehensively assesses models' understanding of physical commonsense. Alongside PhyGenBench, we propose a novel evaluation framework called PhyGenEval. This framework employs a hierarchical evaluation structure utilizing appropriate advanced vision-language models and large language models to assess physical commonsense. Through PhyGenBench and PhyGenEval, we can conduct large-scale automated assessments of T2V models' understanding of physical commonsense, which align closely with human feedback. Our evaluation results and in-depth analysis demonstrate that current models struggle to generate videos that comply with physical commonsense. Moreover, simply scaling up models or employing prompt engineering techniques is insufficient to fully address the challenges presented by PhyGenBench (e.g., dynamic scenarios). We hope this study will inspire the community to prioritize the learning of physical commonsense in these models beyond entertainment applications. We will release the data and codes at https://github.com/OpenGVLab/PhyGenBench

cs.CV↗

arXiv2024

MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models

The capability to process multiple images is crucial for Large Vision-Language Models (LVLMs) to develop a more thorough and nuanced understanding of a scene. Recent multi-image LVLMs have begun to address this need. However, their evaluation has not kept pace with their development. To fill this gap, we introduce the Multimodal Multi-image Understanding (MMIU) benchmark, a comprehensive evaluation suite designed to assess LVLMs across a wide range of multi-image tasks. MMIU encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions, making it the most extensive benchmark of its kind. Our evaluation of 24 popular LVLMs, including both open-source and proprietary models, reveals significant challenges in multi-image comprehension, particularly in tasks involving spatial understanding. Even the most advanced models, such as GPT-4o, achieve only 55.7% accuracy on MMIU. Through multi-faceted analytical experiments, we identify key performance gaps and limitations, providing valuable insights for future model and data improvements. We aim for MMIU to advance the frontier of LVLM research and development, moving us toward achieving sophisticated multimodal multi-image user interactions.

cs.CV↗

arXiv2024

Hilton-Milner theorem for $k$-multisets

Let $ k, n \in \mathbb{N}^+ $ and $ m \in \mathbb{N}^+ \cup \{\infty \} $. A $ k $-multiset in $ [n]_m $ is a $ k $-set whose elements are integers from $ \{1, 2, \ldots, n\} $, and each element is allowed to have at most $ m $ repetitions. A family of $ k $-multisets in $ [n]_m $ is said to be intersecting if every pair of $ k $-multisets from the family have non-empty intersection. In this paper, we give the size and structure of the largest non-trivial intersecting family of $ k $-multisets in $ [n]_m $ for $ n \geq k + \lceil k/m \rceil $. In the special case when $m=\infty$, our result gives rise to an unbounded multiset version for Hilton-Milner Theorem given by Meagher and Purdy. Furthermore, our main theorem unites the statements of the Hilton-Milner Theorem for finite sets and unbounded multisets.

math.CO↗

arXiv2023

Erdős-Ko-Rado Theorem for Bounded Multisets

Let $ k, m, n $ be positive integers with $ k \geq 2 $. A $ k $-multiset of $ [n]_m $ is a collection of $ k $ integers from the set $ \{1, 2, \ldots, n\} $ in which the integers can appear more than once but at most $ m $ times. A family of such $ k $-multisets is called an intersecting family if every pair of $ k $-multisets from the family have non-empty intersection. A finite sequence of real numbers $\{a_1,a_2,\ldots,a_n\}$ is said to be unimodal if there is some $k\in \{1,2,\ldots,n\}$, such that $a_1\leq a_2\leq\ldots\leq a_{k-1}\leq a_k\geq a_{k+1}\geq \ldots\geq a_n$. Given $m,n,k$, denote $C_{k,l}$ as the coefficient of $x^k$ in the generating function $(\sum_{i=1}^mx^i)^l$, where $1\leq l\leq n$. In this paper, we first show that the sequence of $\{C_{k,1},C_{k,2},\ldots,C_{k,n}\}$ is unimodal. Then we use this as a tool to prove that the intersecting family in which every $ k $-multiset contains a fixed element attains the maximum cardinality for $ n \geq k + \lceil k/m\rceil $. In the special case when $m = 1$ and $m=\infty$, our result gives rise to the famous Erdős-Ko-Rado Theorem and an unbounded multiset version for this problem given by Meagher and Purdy, respectively. The main result in this paper can be viewed as a bounded multiset version of the Erdős-Ko-Rado Theorem.

math.CO↗

arXiv2022

Triangles in r-wise t-intersecting families

Let $t$, $r$, $k$ and $n$ be positive integers and $\mathcal{F}$ a family of $k$-subsets of an $n$-set $V$. The family $ \CF $ is $ r $-wise $ t $-intersecting if for any $ F_1, \ldots, F_r \in \CF $, we have $ \abs{\cap_{i = 1}^{r}F_i}\gs t $. An $ r $-wise $ t $-intersecting family of $ r + 1 $ sets $ \{T_1, \ldots, T_{r + 1}\} $ is called an $ (r + 1,t) $-triangle if $ |T_1 \cap \cdots \cap T_{r + 1}| \ls t - 1 $. In this paper, we prove that if $ n \gs n_0(r, t, k) $, then the $ r $-wise $ t $-intersecting family $ \CF \subseteq \binom{[n]}{k} $ containing the most $ (r + 1,t) $-triangles is isomorphic to $ \curlybraces{F \in \binom{[n]}{k}: \abs{F \cap [r + t]} \gs r + t - 1} $. This can also be regarded as a generalized Turán type result.

math.CO↗

arXiv2022

On the $ P_3 $-hull numbers of $ q $-Kneser graphs and Grassmann graphs

Let $S$ be an $n$-dimensional vector space over the finite field $\mathbb{F}_q$, where $q$ is necessarily a prime power. Denote $K_q(n,k)$ (resp. $J_q(n,k)$) to be the \emph{$q$-Kneser graph} (resp. \emph{Grassmann graph}) for $k\geq 1$ whose vertices are the $k$-dimensional subspaces of $S$ and two vertices $v_1$ and $v_2$ are adjacent if $\dim(v_1\cap v_2)=0$ (resp. $\dim(v_1\cap v_2)=k-1$). We consider the infection spreading in the $ q $-Kneser graphs and the Grassmann graphs: a vertex gets infected if it has at least two infected neighbors. In this paper, we compute the $ P_3 $-hull numbers of $K_q(n,k)$ and $J_q(n,k)$ respectively, which is the minimum size of a vertex set that eventually infects the whole graph.

math.CO↗

arXiv2014

A Technology for BigData Analysis Task Description using Domain-Specific Languages

The article presents a technology for dynamic knowledge-based building of Domain-Specific Languages (DSL) to describe data-intensive scientific discovery tasks using BigData technology. The proposed technology supports high level abstract definition of analytic and simulation parts of the task as well as integration into the composite scientific solutions. Automatic translation of the abstract task definition enables seamless integration of various data sources within single solution.

cs.DC↗