Searcharxiv⌕ Search

arXiv subjects

Zhengyu Hu

Publications and source records attributed to Zhengyu Hu.

36 records · Page 2Linked to original sources
arXiv2025

Editable Concept Bottleneck Models

Concept Bottleneck Models (CBMs) have garnered much attention for their ability to elucidate the prediction process through a humanunderstandable concept layer. However, most previous studies focused on cases where the data, including concepts, are clean. In many scenarios, we often need to remove/insert some training data or new concepts from trained CBMs for reasons such as privacy concerns, data mislabelling, spurious concepts, and concept annotation errors. Thus, deriving efficient editable CBMs without retraining from scratch remains a challenge, particularly in large-scale applications. To address these challenges, we propose Editable Concept Bottleneck Models (ECBMs). Specifically, ECBMs support three different levels of data removal: concept-label-level, concept-level, and data-level. ECBMs enjoy mathematically rigorous closed-form approximations derived from influence functions that obviate the need for retraining. Experimental results demonstrate the efficiency and adaptability of our ECBMs, affirming their practical value in CBMs.

cs.LG↗
arXiv2025

LLM-powered Multi-agent Framework for Goal-oriented Learning in Intelligent Tutoring System

Intelligent Tutoring Systems (ITSs) have revolutionized education by offering personalized learning experiences. However, as goal-oriented learning, which emphasizes efficiently achieving specific objectives, becomes increasingly important in professional contexts, existing ITSs often struggle to deliver this type of targeted learning experience. In this paper, we propose GenMentor, an LLM-powered multi-agent framework designed to deliver goal-oriented, personalized learning within ITS. GenMentor begins by accurately mapping learners' goals to required skills using a fine-tuned LLM trained on a custom goal-to-skill dataset. After identifying the skill gap, it schedules an efficient learning path using an evolving optimization approach, driven by a comprehensive and dynamic profile of learners' multifaceted status. Additionally, GenMentor tailors learning content with an exploration-drafting-integration mechanism to align with individual learner needs. Extensive automated and human evaluations demonstrate GenMentor's effectiveness in learning guidance and content quality. Furthermore, we have deployed it in practice and also implemented it as an application. Practical human study with professional learners further highlights its effectiveness in goal alignment and resource targeting, leading to enhanced personalization. Supplementary resources are available at https://github.com/GeminiLight/gen-mentor.

cs.AI↗
arXiv2025

$P$-trivial MMP, Zariski decompositions and minimal models for generalised pairs

We develop a theory of $P$-trivial MMP whose each step is $P$-trivial for a given nef divisor $P$. As an application, we prove that, given a projective generalised klt pair $(X,B+M)$ with data $M'$ being just a nef $\mathbb{R}$-divisor, if $K_X+B+M$ birationally has a Nakayama-Zariski decomposition with nef positive part, and either if $M'$ or the positive part is log numerically effective, then it has a minimal model. Furthermore, we prove this for generalised lc pairs in dimension $3$. This is a generalisation of the main theorem of [Birkar-Hu14]. We also prove some related results.

math.AG↗
arXiv2024

How Many Validation Labels Do You Need? Exploring the Design Space of Label-Efficient Model Ranking

This paper presents LEMR (Label-Efficient Model Ranking) and introduces the MoraBench Benchmark. LEMR is a novel framework that minimizes the need for costly annotations in model selection by strategically annotating instances from an unlabeled validation set. To evaluate LEMR, we leverage the MoraBench Benchmark, a comprehensive collection of model outputs across diverse scenarios. Our extensive evaluation across 23 different NLP tasks in semi-supervised learning, weak supervision, and prompt selection tasks demonstrates LEMR's effectiveness in significantly reducing labeling costs. Key findings highlight the impact of suitable ensemble methods, uncertainty sampling strategies, and model committee selection in enhancing model ranking accuracy. LEMR, supported by the insights from MoraBench, provides a cost-effective and accurate solution for model selection, especially valuable in resource-constrained environments.

cs.LG↗
arXiv2023

UNeR3D: Versatile and Scalable 3D RGB Point Cloud Generation from 2D Images in Unsupervised Reconstruction

In the realm of 3D reconstruction from 2D images, a persisting challenge is to achieve high-precision reconstructions devoid of 3D Ground Truth data reliance. We present UNeR3D, a pioneering unsupervised methodology that sets a new standard for generating detailed 3D reconstructions solely from 2D views. Our model significantly cuts down the training costs tied to supervised approaches and introduces RGB coloration to 3D point clouds, enriching the visual experience. Employing an inverse distance weighting technique for color rendering, UNeR3D ensures seamless color transitions, enhancing visual fidelity. Our model's flexible architecture supports training with any number of views, and uniquely, it is not constrained by the number of views used during training when performing reconstructions. It can infer with an arbitrary count of views during inference, offering unparalleled versatility. Additionally, the model's continuous spatial input domain allows the generation of point clouds at any desired resolution, empowering the creation of high-resolution 3D RGB point clouds. We solidify the reconstruction process with a novel multi-view geometric loss and color loss, demonstrating that our model excels with single-view inputs and beyond, thus reshaping the paradigm of unsupervised learning in 3D vision. Our contributions signal a substantial leap forward in 3D vision, offering new horizons for content creation across diverse applications. Code is available at https://github.com/HongbinLin3589/UNeR3D.

cs.CV↗
arXiv2023

On the Trade-off of Intra-/Inter-class Diversity for Supervised Pre-training

Pre-training datasets are critical for building state-of-the-art machine learning models, motivating rigorous study on their impact on downstream tasks. In this work, we study the impact of the trade-off between the intra-class diversity (the number of samples per class) and the inter-class diversity (the number of classes) of a supervised pre-training dataset. Empirically, we found that with the size of the pre-training dataset fixed, the best downstream performance comes with a balance on the intra-/inter-class diversity. To understand the underlying mechanism, we show theoretically that the downstream performance depends monotonically on both types of diversity. Notably, our theory reveals that the optimal class-to-sample ratio (#classes / #samples per class) is invariant to the size of the pre-training dataset, which motivates an application of predicting the optimal number of pre-training classes. We demonstrate the effectiveness of this application by an improvement of around 2 points on the downstream tasks when using ImageNet as the pre-training dataset.

cs.LG↗
arXiv2023

Laboratory Study of Turbulent Kinetic Energy within Mangrove Forest under Waves

Laboratory study was set out to investigate the flow structure and turbulent kinetic energy ($TKE$) within mangrove forest considering flexible canopies and rigid stems under various wave conditions. Canopies and stems have different effects on turbulence generation. The flexible leaves in canopy scale can promote the dissipation of $TKE$ and the rigid stem scale contribute to the generation of turbulence. Importantly, the difference of $TKE$ with and without vegetation, $ΔTKE$ varied with wave frequency and peaked at 0.833 Hz for the canopy and vegetation-free upper layer, but increased with wave frequency in stem scale. Moreover, $TKE$ in mangrove forest increased monotonically with ${u_{w,RMS}}^2$ for each stem density. This study proposed a modified model for the $TKE$ prediction within canopy scale considering the role of flexible leaves and rigid stems together. Besides, $TKE/{u_{w,RMS}}^2$ presented a good correlation with the ratio of wave excursion (${A_w}$) to stem spacing ($S$), ${A_w}$/$S$ and declined rapidly with ${A_w}$/$S$ especially when ${A_w}$/$S$ <0.1 for both stem and canopy scale. This empirical relationship can be a predictor for $TKE$ within mangrove forest. Further, considering the scale between wave energy dissipation by vegetation and production of $TKE$, the third model was validated in mangrove forest and predicted well.

physics.flu-dyn↗
arXiv2023

Laboratory Study of Wave Attenuation by Mangrove Forest

Coastal vegetation can dissipate the energy of storm surge and tsunami, which serves a natural barrier for coastal protection. The experimental study of waves-mangroves interaction was carried out to evaluate the efficiency of wave attenuation by vegetation (WDV). In particular, the model mangrove consisted of a flexible canopy and a rigid stem to imitate natural Rhizophora better. It considered the effect of various wave conditions including incident wave height (${H_i}$) and wave period ($T$) as well as plant characteristics including submergence ratio ($α$) and stem density ($N$), respectively. Then, the exponential wave decay coefficient (${K_i}$) was applied to determine the performance of WDV. It showed that larger ${H_i}$, $α$ and smaller $T$ induced greater wave decay. To quantify WDV, a new analytical model was proposed considering the change of stem arrangement and wave parameters along the vegetation. Further, the drag coefficient, ${C_D}$ of mangrove forest was analytical calibrated by this model, and made a comparison with the previous model. Based on this model, the relationships between ${C_D}$ and some non-dimensional parameters, e.g. Reynolds number, ${\mathop{\rm Re}\nolimits} $ , Keulegan-Carpenter number, $KC$ and Ursell number, ${U_r}$ were established to evaluate and predict the wave decay. Especially, it showed a better correlation relationship than the previous model, and the predicted results showed a good agreement with the experimental data.

physics.flu-dyn↗
arXiv2023

Experimental study on wave attenuation by vegetation by means of particle imaging velocimetry

Coastal vegetation serves as an ecological defense and plays an important role when facing marine disaster such as storm surge. In this study, we investigated experimentally the wave attenuation, including wave height and wave force decay, and flow field within rigid vegetation by means of particle imaging velocimetry (PIV). Drag coefficient of vegetation is closely related the wave attenuation by vegetation. Based on Morison equation, a new modified direct measurement method to derive the drag coefficient of vegetation was proposed in this study. This method can eliminate the potential errors and obtain more accurate drag coefficient to have a well-fitted relationship with Reynolds number compared with the previous methods. More importantly, we showed the flow field including instantaneous velocity filed in different wave phase within the vegetation and turbulent kinetic energy (TKE) field to describe the mechanisms of wave energy dissipation. In addition, a prediction model to reproduce the instantaneous TKE was proposed by the work input of vegetation. The results illustrated that all these characteristics of parameters can help us understand the effect of the coastal vegetation on wave attenuation better.

physics.flu-dyn↗
arXiv2021

An abundance theorem for generalised pairs

We prove the finiteness of $B$-representations of generalised log canonical pairs. As a consequence, we prove that, the (relative) abundance for a generalised semi-log canonical pair is implied by the abundance for its normalisation. Furthermore, we obtain an abundance theorem for generalised dlt pairs with abundant data, which is a natural inductive step of generalised minimal model program.

math.AG↗
arXiv2021

Log abundance of the moduli b-divisors of lc-trivial fibrations

We prove that the moduli b-divisor of an lc-trivial fibration from a log canonical pair is log abundant. The result follows from a theorem on the restriction of the moduli b-divisor, based on a theory of lc-trivial morphisms, which allows us to treat $\mathbb{R}$-divisors and proper morphisms possibly with disconnected fibres. We also prove a theorem on extending a finite cover over a closed subvariety to that over a variety in arbitrary characteristic.

math.AG↗
arXiv2019

On minimal model theory for log abundant lc pairs

Under the assumption of the minimal model theory for projective klt pairs of dimension $n$, we establish the minimal model theory for lc pairs $(X/Z,Δ)$ such that the log canonical divisor is relatively log abundant and its restriction to any lc center has relative numerical dimension at most $n$. We also give another detailed proof of results by the second author, and study termination of log MMP with scaling.

math.AG↗
arXiv2019

Log canonical pairs with boundaries containing ample divisors

Let $(X,Δ)$ be a projective log canonical pair such that $Δ\geq A$ where $A \geq 0$ is an ample $\mathbb{R}$-divisor. We prove that either $(X,Δ)$ has a good minimal model or a Mori fibre space. Moreover, if $X$ is $\mathbb{Q}$-factorial, then any Log Minimal Model Program on $K_X+Δ$ with scaling terminates. As an application we prove that a log Fano type variety $X$ with $\mathbb{Q}$-factorial log canonical singularities is a Mori dream space.

math.AG↗
arXiv2018

Log canonical pairs over varieties with maximal Albanese dimension

Let $(X,B)$ be a log canonical pair over a normal variety $Z$ with maximal Albanese dimension. If $K_X+B$ is relatively abundant over $Z$ (for example, $K_X+B$ is relatively big over $Z$), then we prove that $K_X+B$ is abundant. In particular, the subadditvity of Kodaira dimensions $κ(K_X+B) \geq κ(K_F+B_F)+ κ(Z)$ holds, where $F$ is a general fiber, $K_F+B_F= (K_X+B)|_F$, and $κ(Z)$ means the Kodaira dimension of a smooth model of $Z$. We discuss several variants of this result in Section 4. We also give a remark on the log Iitaka conjecture for log canonical pairs in Section 5.

math.AG↗
arXiv2014

Valuative multiplier ideals

The main goal of this paper is to construct an algebraic analogue of quasi-plurisubharmonic function (qpsh for short) from complex analysis and geometry. We define a notion of qpsh function on a valuation space associated to a quite general scheme. We then define the multiplier ideals of these functions and prove some basic results about them, such as subadditivity property, the approximation theorem. We also treat some applications in complex algebraic geometry.

math.AG↗
arXiv2013

Log canonical pairs with good augmented base loci

Let $(X,B)$ be a projective log canonical pair such that $B$ is a $\Q$-divisor, and that there is a surjective morphism $f\colon X\to Z$ onto a normal variety $Z$ satisfying: $K_X+B\sim_\Q f^*M$ for some $\Q$-divisor $M$, and the augmented base locus ${\bf{B_+}}(M)$ does not contain the image of any log canonical centre of $(X,B)$. We will show that $(X,B)$ has a good log minimal model. An interesting special case is when $f$ is the identity morphism.

math.AG↗
arXiv2013

On log minimal models and Zariski decompositions II

We continue our study of the relation between log minimal models and various types of Zariski decompositions. Let $(X,B)$ be a projective log canonical pair. We will show that $(X,B)$ has a log minimal model if either $K_X+B$ birationally has a Nakayama-Zariski decomposition with nef positive part, or that $K_X+B$ is big and birationally it has a Fujita or CKM Zariski decomposition. Along the way we introduce polarized pairs $(X,B+P)$ where $(X,B)$ is a usual projective pair and $P$ is nef, and study the birational geometry of such pairs.

math.AG↗