arXiv subjects

Zijian Zhang

Publications and source records attributed to Zijian Zhang.

At least 163 records · Page 9Linked to original sources

arXiv2025

GRAPE: Generalizing Robot Policy via Preference Alignment

Despite the recent advancements of vision-language-action (VLA) models on a variety of robotics tasks, they suffer from critical issues such as poor generalizability to unseen tasks, due to their reliance on behavior cloning exclusively from successful rollouts. Furthermore, they are typically fine-tuned to replicate demonstrations collected by experts under different settings, thus introducing distribution bias and limiting their adaptability to diverse manipulation objectives, such as efficiency, safety, and task completion. To bridge this gap, we introduce GRAPE: Generalizing Robot Policy via Preference Alignment. Specifically, GRAPE aligns VLAs on a trajectory level and implicitly models reward from both successful and failure trials to boost generalizability to diverse tasks. Moreover, GRAPE breaks down complex manipulation tasks to independent stages and automatically guides preference modeling through customized spatiotemporal constraints with keypoints proposed by a large vision-language model. Notably, these constraints are flexible and can be customized to align the model with varying objectives, such as safety, efficiency, or task success. We evaluate GRAPE across a diverse array of tasks in both real-world and simulated environments. Experimental results demonstrate that GRAPE enhances the performance of state-of-the-art VLA models, increasing success rates on in-domain and unseen manipulation tasks by 51.79% and 58.20%, respectively. Additionally, GRAPE can be aligned with various objectives, such as safety and efficiency, reducing collision rates by 37.44% and rollout step-length by 11.15%, respectively. All code, models, and data are available at https://grape-vla.github.io/

cs.RO↗

arXiv2025

Large Language Model Distilling Medication Recommendation Model

The recommendation of medication is a vital aspect of intelligent healthcare systems, as it involves prescribing the most suitable drugs based on a patient's specific health needs. Unfortunately, many sophisticated models currently in use tend to overlook the nuanced semantics of medical data, while only relying heavily on identities. Furthermore, these models face significant challenges in handling cases involving patients who are visiting the hospital for the first time, as they lack prior prescription histories to draw upon. To tackle these issues, we harness the powerful semantic comprehension and input-agnostic characteristics of Large Language Models (LLMs). Our research aims to transform existing medication recommendation methodologies using LLMs. In this paper, we introduce a novel approach called Large Language Model Distilling Medication Recommendation (LEADER). We begin by creating appropriate prompt templates that enable LLMs to suggest medications effectively. However, the straightforward integration of LLMs into recommender systems leads to an out-of-corpus issue specific to drugs. We handle it by adapting the LLMs with a novel output layer and a refined tuning loss function. Although LLM-based models exhibit remarkable capabilities, they are plagued by high computational costs during inference, which is impractical for the healthcare sector. To mitigate this, we have developed a feature-level knowledge distillation technique, which transfers the LLM's proficiency to a more compact model. Extensive experiments conducted on two real-world datasets, MIMIC-III and MIMIC-IV, demonstrate that our proposed model not only delivers effective results but also is efficient. To ease the reproducibility of our experiments, we release the implementation code online.

cs.IR↗

arXiv2025

Multi-source Multi-level Multi-token Ethereum Dataset and Benchmark Platform

This paper introduces 3MEthTaskforce (https://3meth.github.io), a multi-source, multi-level, and multi-token Ethereum dataset addressing the limitations of single-source datasets. Integrating over 300 million transaction records, 3,880 token profiles, global market indicators, and Reddit sentiment data from 2014-2024, it enables comprehensive studies on user behavior, market sentiment, and token performance. 3MEthTaskforce defines benchmarks for user behavior prediction and token price prediction tasks, using 6 dynamic graph networks and 19 time-series models to evaluate performance. Its multimodal design supports risk analysis and market fluctuation modeling, providing a valuable resource for advancing blockchain analytics and decentralized finance research.

cs.CE↗

arXiv2025

Compact Model of Linear Passive Integrated Photonics Device for Photon Design Automation

As integrated photonic systems grow in scale and complexity, Photonic Design Automation (PDA) tools and Process Design Kits (PDKs) have become increasingly important for layout and simulation. However, fixed PDKs often fail to meet the rising demand for customization, compelling designers to spend significant time on geometry optimization using FDTD, EME, and BPM simulations. To address this challenge, we propose a data-driven Eigenmode Propagation Method (DEPM) based on the unitary evolution of optical waveguides, along with a compact model derived from intrinsic waveguide Hamiltonians. The relevant parameters are extracted via complex coupled-mode theory. Once constructed, the compact model enables millisecond-scale simulations that achieve accuracy on par with 3D FDTD, within the model's valid scope. Moreover, this method can swiftly evaluate the effects of manufacturing variations on device and system performance, including both random phase errors and polarization-sensitive components. The data-driven EPM thus provides an efficient and flexible solution for future photonic design automation, promising further advancements in integrated photonic technologies.

physics.optics↗

arXiv2025

Nonreciprocal Optical Routing in Multi-port Magneto-Optical Devices on Silicon

Nonreciprocal optical devices are key components in photonic integrated circuits for light reflection blocking and routing. Most reported silicon integrated nonreciprocal optical devices to date were unit devices. To allow complex signal routing between multi-ports in photonic networks, multi-port magneto-optical (MO) nonreciprocal photonic devices are desired. In this study, we report experimental demonstration of a silicon integrated 5*5 multiport nonreciprocal photonic device based on magneto-optical waveguides. By introducing different nonreciprocal phase shift effect to planar photonic waveguides, the device focuses light to different ports for both forward and backward propagation. The device shows designable nonreciprocal transmission between 5*5 ports, achieving 16 dB isolation ratio and -18 dB crosstalk.

physics.optics↗

arXiv2024

A Contrastive Pretrain Model with Prompt Tuning for Multi-center Medication Recommendation

Medication recommendation is one of the most critical health-related applications, which has attracted extensive research interest recently. Most existing works focus on a single hospital with abundant medical data. However, many small hospitals only have a few records, which hinders applying existing medication recommendation works to the real world. Thus, we seek to explore a more practical setting, i.e., multi-center medication recommendation. In this setting, most hospitals have few records, but the total number of records is large. Though small hospitals may benefit from total affluent records, it is also faced with the challenge that the data distributions between various hospitals are much different. In this work, we introduce a novel conTrastive prEtrain Model with Prompt Tuning (TEMPT) for multi-center medication recommendation, which includes two stages of pretraining and finetuning. We first design two self-supervised tasks for the pretraining stage to learn general medical knowledge. They are mask prediction and contrastive tasks, which extract the intra- and inter-relationships of input diagnosis and procedures. Furthermore, we devise a novel prompt tuning method to capture the specific information of each hospital rather than adopting the common finetuning. On the one hand, the proposed prompt tuning can better learn the heterogeneity of each hospital to fit various distributions. On the other hand, it can also relieve the catastrophic forgetting problem of finetuning. To validate the proposed model, we conduct extensive experiments on the public eICU, a multi-center medical dataset. The experimental results illustrate the effectiveness of our model. The implementation code is available to ease the reproducibility https://github.com/Applied-Machine-Learning-Lab/TEMPT.

cs.IR↗

arXiv2024

Focusing Image Generation to Mitigate Spurious Correlations

Instance features in images exhibit spurious correlations with background features, affecting the training process of deep neural classifiers. This leads to insufficient attention to instance features by the classifier, resulting in erroneous classification outcomes. In this paper, we propose a data augmentation method called Spurious Correlations Guided Synthesis (SCGS) that mitigates spurious correlations through image generation model. This approach does not require expensive spurious attribute (group) labels for the training data and can be widely applied to other debiasing methods. Specifically, SCGS first identifies the incorrect attention regions of a pre-trained classifier on the training images, and then uses an image generation model to generate new training data based on these incorrect attended regions. SCGS increases the diversity and scale of the dataset to reduce the impact of spurious correlations on classifiers. Changes in the classifier's attention regions and experimental results on three different domain datasets demonstrate that this method is effective in reducing the classifier's reliance on spurious correlations.

cs.CV↗

arXiv2024

Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach

Multifaceted user modeling aims to uncover fine-grained patterns and learn representations from user data, revealing their diverse interests and characteristics, such as profile, preference, and personality. Recent studies on foundation model-based recommendation have emphasized the Transformer architecture's remarkable ability to capture complex, non-linear user-item interaction relationships. This paper aims to advance foundation model-based recommendersystems by introducing enhancements to multifaceted user modeling capabilities. We propose a novel Transformer layer designed specifically for recommendation, using the self-attention mechanism to capture sequential user-item interaction patterns. Specifically, we design a group gating network to identify user groups, enabling hierarchical discovery across different layers, thereby capturing the multifaceted nature of user interests through multiple Transformer layers. Furthermore, to broaden the data scope and further enhance multifaceted user modeling, we extend the framework to a federated setting, enabling the use of private datasets while ensuring privacy. Experimental validations on benchmark datasets demonstrate the superior performance of our proposed method. Code is available.

cs.IR↗

arXiv2024

LLM-Powered User Simulator for Recommender System

User simulators can rapidly generate a large volume of timely user behavior data, providing a testing platform for reinforcement learning-based recommender systems, thus accelerating their iteration and optimization. However, prevalent user simulators generally suffer from significant limitations, including the opacity of user preference modeling and the incapability of evaluating simulation accuracy. In this paper, we introduce an LLM-powered user simulator to simulate user engagement with items in an explicit manner, thereby enhancing the efficiency and effectiveness of reinforcement learning-based recommender systems training. Specifically, we identify the explicit logic of user preferences, leverage LLMs to analyze item characteristics and distill user sentiments, and design a logical model to imitate real human engagement. By integrating a statistical model, we further enhance the reliability of the simulation, proposing an ensemble model that synergizes logical and statistical insights for user interaction simulations. Capitalizing on the extensive knowledge and semantic generation capabilities of LLMs, our user simulator faithfully emulates user behaviors and preferences, yielding high-fidelity training data that enrich the training of recommendation algorithms. We establish quantifying and qualifying experiments on five datasets to validate the simulator's effectiveness and stability across various recommendation scenarios.

cs.IR↗

arXiv2024

DNS-Rec: Data-aware Neural Architecture Search for Recommender Systems

In the era of data proliferation, efficiently sifting through vast information to extract meaningful insights has become increasingly crucial. This paper addresses the computational overhead and resource inefficiency prevalent in existing Sequential Recommender Systems (SRSs). We introduce an innovative approach combining pruning methods with advanced model designs. Furthermore, we delve into resource-constrained Neural Architecture Search (NAS), an emerging technique in recommender systems, to optimize models in terms of FLOPs, latency, and energy consumption while maintaining or enhancing accuracy. Our principal contribution is the development of a Data-aware Neural Architecture Search for Recommender System (DNS-Rec). DNS-Rec is specifically designed to tailor compact network architectures for attention-based SRS models, thereby ensuring accuracy retention. It incorporates data-aware gates to enhance the performance of the recommendation network by learning information from historical user-item interactions. Moreover, DNS-Rec employs a dynamic resource constraint strategy, stabilizing the search process and yielding more suitable architectural solutions. We demonstrate the effectiveness of our approach through rigorous experiments conducted on three benchmark datasets, which highlight the superiority of DNS-Rec in SRSs. Our findings set a new standard for future research in efficient and accurate recommendation systems, marking a significant step forward in this rapidly evolving field.

cs.IR↗

arXiv2024

GARLIC: GPT-Augmented Reinforcement Learning with Intelligent Control for Vehicle Dispatching

As urban residents demand higher travel quality, vehicle dispatch has become a critical component of online ride-hailing services. However, current vehicle dispatch systems struggle to navigate the complexities of urban traffic dynamics, including unpredictable traffic conditions, diverse driver behaviors, and fluctuating supply and demand patterns. These challenges have resulted in travel difficulties for passengers in certain areas, while many drivers in other areas are unable to secure orders, leading to a decline in the overall quality of urban transportation services. To address these issues, this paper introduces GARLIC: a framework of GPT-Augmented Reinforcement Learning with Intelligent Control for vehicle dispatching. GARLIC utilizes multiview graphs to capture hierarchical traffic states, and learns a dynamic reward function that accounts for individual driving behaviors. The framework further integrates a GPT model trained with a custom loss function to enable high-precision predictions and optimize dispatching policies in real-world scenarios. Experiments conducted on two real-world datasets demonstrate that GARLIC effectively aligns with driver behaviors while reducing the empty load rate of vehicles.

cs.LG↗

arXiv2024

Modeling Multiple X-Ray Reflection in Super-Eddington Winds

It has been recently discovered that a few super-Eddington sources undergoing black hole super-Eddington accretion exhibit X-ray reflection signatures. In such new systems, one expects that the coronal X-ray emissions are mainly reflected by optically thick super-Eddington winds instead of thin disks. In this paper, we conduct a series of general relativistic ray-tracing and Monte Carlo radiative transfer simulations to model the X-ray reflection signatures, especially the characteristic Fe K$α$ line, produced from super-Eddington accretion flows around non-spinning black holes. In particular, we allow the photons emitted by a lamppost corona to be reflected multiple times in a cone-like funnel surrounded by fast winds. We find that the Fe K$α$ line profile most sensitively depends on the wind kinematics, while its exact shape also depends on the funnel open angle and corona height. Furthermore, very interestingly, we find that the Fe K$α$ line can have a prominent double-peak profile in certain parameter spaces even with a face-on orientation. Moreover, we compare the Fe K$α$ line profiles produced from super-Eddington and thin disks and show that such lines can provide important insights into the understanding of black hole systems undergoing super-Eddington accretion.

astro-ph.HE↗

arXiv2024

A Comparative Evaluation of Automated Analysis Tools for Solidity Smart Contracts

Blockchain smart contracts have emerged as a transformative force in the digital realm, spawning a diverse range of compelling applications. Since solidity smart contracts across various domains manage trillions of dollars in virtual coins, they become a prime target for attacks. One of the primary challenges is keeping abreast of the latest techniques and tools for developing secure smart contracts and examining those already deployed. In this paper, we seek to address these challenges from four aspects: (1) We begin by examining ten automatic tools, specifically focusing on their methodologies and their ability to identify vulnerabilities in solidity smart contracts. (2) We propose a novel criterion for evaluating these tools, based on the ISO/IEC 25010 standard. (3) To facilitate the evaluation of the selected tools, we construct a benchmark that encompasses two distinct datasets: a collection of 389 labelled smart contracts and a scaled set of 20,000 unique cases from real-world contracts. (4) We provide a comparison of the selected tools, offering insights into their strengths and weaknesses and highlighting areas where further improvements are needed. Through this evaluation, we hope to provide developers and researchers with valuable guidance on selecting and using smart contract analysis tools and contribute to the ongoing efforts to improve the security and reliability of smart contracts.

cs.DC↗

arXiv2024

Successive Bayesian Reconstructor for Channel Estimation in Fluid Antenna Systems

Fluid antenna systems (FASs) can reconfigure their antenna locations freely within a spatially continuous space. To keep favorable antenna positions, the channel state information (CSI) acquisition for FASs is essential. While some techniques have been proposed, most existing FAS channel estimators require several channel assumptions, such as slow variation and angular-domain sparsity. When these assumptions are not reasonable, the model mismatch may lead to unpredictable performance losses. In this paper, we propose the successive Bayesian reconstructor (S-BAR) as a general solution to estimate FAS channels. Unlike model-based estimators, the proposed S-BAR is prior-aided, which builds the experiential kernel for CSI acquisition. Inspired by Bayesian regression, the key idea of S-BAR is to model the FAS channels as a stochastic process, whose uncertainty can be successively eliminated by kernel-based sampling and regression. In this way, the predictive mean of the regressed stochastic process can be viewed as a Bayesian channel estimator. Simulation results verify that, in both model-mismatched and model-matched cases, the proposed S-BAR can achieve higher estimation accuracy than the existing schemes.

cs.IT↗

arXiv2024

Photometric Selection of type 1 Quasars in the XMM-LSS Field with Machine Learning and the Disk-Corona Connection

We present photometric selection of type 1 quasars in the $\approx5.3~{\rm deg}^{2}$ XMM-Large Scale Structure (XMM-LSS) survey field with machine learning. We constructed our training and \hbox{blind-test} samples using spectroscopically identified SDSS quasars, galaxies, and stars. We utilized the XGBoost machine learning method to select a total of 1\,591 quasars. We assessed the classification performance based on the blind-test sample, and the outcome was favorable, demonstrating high reliability ($\approx99.9\%$) and good completeness ($\approx87.5\%$). We used XGBoost to estimate photometric redshifts of our selected quasars. The estimated photometric redshifts span a range from 0.41 to 3.75. The outlier fraction of these photometric redshift estimates is $\approx17\%$ and the normalized median absolute deviation ($σ_{\rm NMAD}$) is $\approx0.07$. To study the quasar disk-corona connection, we constructed a subsample of 1\,016 quasars with HSC $i<22.5$ after excluding radio-loud and potentially X-ray-absorbed quasars. The relation between the optical-to-X-ray power-law slope parameter ($α_{\rm OX}$) and the 2500 Angstrom monochromatic luminosity ($L_{2500}$) for this subsample is $α_{\rm OX}=(-0.156\pm0.007)~{\rm log}~{L_{\rm 2500}}+(3.175\pm0.211)$ with a dispersion of 0.159. We found this correlation in good agreement with the correlations in previous studies. We explored several factors which may bias the $α_{\rm OX}$-$L_{\rm 2500}$ relation and found that their effects are not significant. We discussed possible evolution of the $α_{\rm OX}$-$L_{\rm 2500}$ relation with respect to $L_{\rm 2500}$ or redshift.

astro-ph.GA↗

arXiv2024

Image Understanding Makes for A Good Tokenizer for Image Generation

Abstract Modern image generation (IG) models have been shown to capture rich semantics valuable for image understanding (IU) tasks. However, the potential of IU models to improve IG performance remains uncharted. We address this issue using a token-based IG framework, which relies on effective tokenizers to project images into token sequences. Currently, pixel reconstruction (e.g., VQGAN) dominates the training objective for image tokenizers. In contrast, our approach adopts the feature reconstruction objective, where tokenizers are trained by distilling knowledge from pretrained IU encoders. Comprehensive comparisons indicate that tokenizers with strong IU capabilities achieve superior IG performance across a variety of metrics, datasets, tasks, and proposal networks. Notably, VQ-KD CLIP achieves $4.10$ FID on ImageNet-1k (IN-1k). Visualization suggests that the superiority of VQ-KD can be partly attributed to the rich semantics within the VQ-KD codebook. We further introduce a straightforward pipeline to directly transform IU encoders into tokenizers, demonstrating exceptional effectiveness for IG tasks. These discoveries may energize further exploration into image tokenizer research and inspire the community to reassess the relationship between IU and IG. The code is released at https://github.com/magic-research/vector_quantization.

cs.CV↗

arXiv2024

DISCO: DISCovering Overfittings as Causal Rules for Text Classification Models

With the rapid advancement of neural language models, the deployment of over-parameterized models has surged, increasing the need for interpretable explanations comprehensible to human inspectors. Existing post-hoc interpretability methods, which often focus on unigram features of single input textual instances, fail to capture the models' decision-making process fully. Additionally, many methods do not differentiate between decisions based on spurious correlations and those based on a holistic understanding of the input. Our paper introduces DISCO, a novel method for discovering global, rule-based explanations by identifying causal n-gram associations with model predictions. This method employs a scalable sequence mining technique to extract relevant text spans from training data, associate them with model predictions, and conduct causality checks to distill robust rules that elucidate model behavior. These rules expose potential overfitting and provide insights into misleading feature combinations. We validate DISCO through extensive testing, demonstrating its superiority over existing methods in offering comprehensive insights into complex model behaviors. Our approach successfully identifies all shortcuts manually introduced into the training data (100% detection rate on the MultiRC dataset), resulting in an 18.8% regression in model performance -- a capability unmatched by any other method. Furthermore, DISCO supports interactive explanations, enabling human inspectors to distinguish spurious causes in the rule-based output. This alleviates the burden of abundant instance-wise explanations and helps assess the model's risk when encountering out-of-distribution (OOD) data.

cs.AI↗

arXiv2024

Efficient and Robust Regularized Federated Recommendation

Recommender systems play a pivotal role across practical scenarios, showcasing remarkable capabilities in user preference modeling. However, the centralized learning paradigm predominantly used raises serious privacy concerns. The federated recommender system (FedRS) addresses this by updating models on clients, while a central server orchestrates training without accessing private data. Existing FedRS approaches, however, face unresolved challenges, including non-convex optimization, vulnerability, potential privacy leakage risk, and communication inefficiency. This paper addresses these challenges by reformulating the federated recommendation problem as a convex optimization issue, ensuring convergence to the global optimum. Based on this, we devise a novel method, RFRec, to tackle this optimization problem efficiently. In addition, we propose RFRecF, a highly efficient version that incorporates non-uniform stochastic gradient descent to improve communication efficiency. In user preference modeling, both methods learn local and global models, collaboratively learning users' common and personalized interests under the federated learning setting. Moreover, both methods significantly enhance communication efficiency, robustness, and privacy protection, with theoretical support. Comprehensive evaluations on four benchmark datasets demonstrate RFRec and RFRecF's superior performance compared to diverse baselines.

cs.IR↗