Yang Shi, Yifeng Xie, Minzhe Guo, Liangsi Lu, Mingxuan Huang, Jingchao Wang, Zhihong Zhu, Boyan Xu, Zhiqi Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In this work, we investigate an intriguing and prevalent phenomenon of diffusion models which we term as "consistent model reproducibility'': given the same starting noise input and a deterministic sampler, different diffusion models often yield remarkably similar outputs. We confirm this phenomenon through comprehensive experiments, implying that different diffusion models consistently reach the same data distribution and score function regardless of diffusion model frameworks, model architectures, or training procedures. More strikingly, our further investigation implies that diffusion models are learning *distinct distributions* influenced by the training data size. This is evident in two distinct training regimes: (I) "memorization regime,'' where the diffusion model overfits to the training data distribution, and (ii) "generalization regime,'' where the model learns the underlying data distribution. Our study also finds that this valuable property generalizes to many variants of diffusion models, including those for conditional generation and solving inverse problems. Lastly, we discuss how our findings connect to existing research and highlight the practical implications of our discoveries.

Recently, diffusion models have emerged as a powerful class of generative models. Despite their success, there is still limited understanding of their semantic spaces. This makes it challenging to achieve precise and disentangled image generation without additional training, especially in an unsupervised way. In this work, we improve the understanding of their semantic spaces from intriguing observations: among a certain range of noise levels, (1) the learned posterior mean predictor (PMP) in the diffusion model is locally linear, and (2) the singular vectors of its Jacobian lie in low-dimensional semantic subspaces. We provide a solid theoretical basis to justify the linearity and low-rankness in the PMP. These insights allow us to propose an unsupervised, single-step, training-free **LO**w-rank **CO**ntrollable image editing (LOCO Edit) method for precise local editing in diffusion models. LOCO Edit identified editing directions with nice properties: homogeneity, transferability, composability, and linearity. These properties of LOCO Edit benefit greatly from the low-dimensional semantic subspace. Our method can further be extended to unsupervised or text-supervised editing in various text-to-image diffusion models (T-LOCO Edit). Finally, extensive empirical experiments demonstrate the effectiveness and efficiency of LOCO Edit. The code and the arXiv version can be found on the [project website](https://chicychen.github.io/LOCO).

Despite the considerable success of neural networks in security settings such as malware detection, such models have proved vulnerable to evasion attacks, in which attackers make slight changes to inputs (e.g., malware) to bypass detection. We propose a novel approach, Fourier stabilization, for designing evasion-robust neural networks with binary inputs. This approach, which is complementary to other forms of defense, replaces the weights of individual neurons with robust analogs derived using Fourier analytic tools. The choice of which neurons to stabilize in a neural network is then a combinatorial optimization problem, and we propose several methods for approximately solving it. We provide a formal bound on the per-neuron drop in accuracy due to Fourier stabilization, and experimentally demonstrate the effectiveness of the proposed approach in boosting robustness of neural networks in several detection settings. Moreover, we show that our approach effectively composes with adversarial training.

Replica placement (RP) is an important technology in content delivery systems to achieve efficient and effective content delivery. Peer-assisted content delivery network (PCDN) has been proposed for simultaneously obtaining the scalability advantage of peer-to-peer networks and the reliability and manageability advantages of content delivery networks. However, the benefits of peer assistance can be severely affected by the self-interested or malicious behaviors of peers. In this research, we utilize mechanism design theory to design robust algorithms for RP in PCDN with the coexistence of altruistic, self-interested, and malicious seeders. Our preliminary results show that RP algorithms that consider peer contributions have better performance in PCDN, in addition, our mechanism incentivizes the contribution of self-interested seeders and further improves the performance of RP in PCDN. We also quantify the impact of malicious behavior on the performance of the mechanism and extend the mechanism to tolerate malicious behaviors.

Research on effective and efficient mobile threat analysis becomes an emerging and important topic in cybersecurity research area. Static analysis and dynamic analysis constitute two of the most popular types of techniques for security analysis and evaluation; nevertheless, each of them has its strengths and weaknesses. To leverage the benefits of both approaches, we propose a hybrid approach that integrates the static and dynamic analysis for detecting security threats in mobile applications. The key of this approach is the unification of data states and software execution on critical test paths. The approach consists of two phases. In the first phase, a pilot static analysis is conducted to identify potential critical attack paths based on Android APIs and existing attack patterns. In the second phase, a dynamic analysis follows the identified critical paths to execute the program in a limited and focused manner. Attacks shall be detected by checking the conformance of the detected paths with existing attack patterns. The method will report the types of detected attack scenarios based on types of sensitive data that may be compromised, such as web browser cookie.

A major challenge in developmental biology is to understand the genetic and cellular processes/programs driving organ formation and differentiation of the diverse cell types that comprise the embryo. While recent studies using single cell transcriptome analysis illustrate the power to measure and understand cellular heterogeneity in complex biological systems, processing large amounts of RNA-seq data from heterogeneous cell populations creates the need for readily accessible tools for the analysis of single-cell RNA-seq (scRNA-seq) profiles. The present study presents a generally applicable analytic pipeline (SINCERA: a computational pipeline for SINgle CEll RNA-seq profiling Analysis) for processing scRNA-seq data from a whole organ or sorted cells. The pipeline supports the analysis for: 1) the distinction and identification of major cell types; 2) the identification of cell type specific gene signatures; and 3) the determination of driving forces of given cell types. We applied this pipeline to the RNA-seq analysis of single cells isolated from embryonic mouse lung at E16.5. Through the pipeline analysis, we distinguished major cell types of fetal mouse lung, including epithelial, endothelial, smooth muscle, pericyte, and fibroblast-like cell types, and identified cell type specific gene signatures, bioprocesses, and key regulators. SINCERA is implemented in R, licensed under the GNU General Public License v3, and freely available from CCHMC PBGE website, https://research.cchmc.org/pbge/sincera.html.

To enhance the Information Assurance and Security in computer science (CS) education and to prepare students for the challenges of security issues, this paper presents our ongoing work on developing mobile-based hands-on security learning modules that will be integrated into existing courses in CS curriculum.

This work-in-progress paper presents an approach to authentic learning of mobile security through real-world-scenario case studies. Five sets of case studies are being developed to cover the state-of-the-art of mobile security knowledge and practices. Some of the developed case studies are being implemented in related courses and the preliminary feedback is positive.

To overcome shortcomings in existing embedded system education, this paper presents an ongoing work on developing a novel Android-based Mobile Sensory System Lab ware that is portable, affordable, and easily adoptable. The proposed lab ware can facilitate authentic and creative learning of embedded systems and provide extensive hands-on experience using mobile embedded platforms. Preliminary feedback from students is positive and encouraging.

Computer networking/security courses are essential parts of college computer science curriculum, and hands-on networking/security experience is well recognized as an effective approach to help students understand better about the network concepts, the layered architecture of network protocols, and network security. A new affordable and easily-adoptable approach to networking/security labs is desirable to enhance network teaching and learning. In this paper, we present our mobile device-based hands-on labware for computer network/security learning, which is guided by authentic learning principles to immerse students in a real world relevant learning environment. The developed labware has been used in various networking and security related courses, and students claim that they can learn more with the hands-on authentic learning experience.

As smart mobile devices grow increasingly in popularity, so do the incentives for attackers. Recent surveys on mobile security describe the rapidly increasing number and sophistication of mobile attacks. Newer sources of risks are being introduced or explored in the mobile computing paradigm where traditional security threats are also evolving. The prevalence of mobile devices and the rapid growth of mobile threats have resulted in a shortage of mobile security personnel. Educational activities are needed to promote mobile security education and to meet the emerging industry and education needs. This paper presents our initial effort on exploring a learning approach to mobile security, which aims at taking advantages of the benefits of mobile devices and the best practices in learning information security, promoting students' interests, and improving students' self-efficacy. An Android security labware is designed to implement the environment and materials for the learning approach. We integrated the pilot modules of the labware into two security courses in two semesters. The majority of the students provided positive feedback and enjoyed the Android security practices.

Digital mobile devices, such as smart phones, have become an essential part of daily lives for the young generation, especially college students. The rapid growth of wireless and mobile technologies results in a huge demand on qualified professionals in this area. More and more schools started teaching wireless and mobile networking, but how to provide students with hands-on practice experience is still a challenge. The networking capabilities of smart mobile devices make it a perfect instructional tool for networking related course teaching. In this study, we have designed and developed a labware with Android smart phones that focuses on teaching and learning of fundamental data communications and networking concepts by making use of the mobile nature of smart phones and mobile devices. The goal of this labware is to provide students with hands-on experience with mobile devices to better understand networking concepts such as different layers of TCP/IP protocol, Bluetooth, Wi-Fi, and a wide range of networking methodologies/implementations in a easy-adoptable, affordable, and effective way.

To address the need for innovative mobile security learning materials and for promoting mobile security education, this paper presents our work-in-progress effort on developing a real-world relevant security labware to provide students with mobile threat analysis and protection experience. A preliminary evaluation has been conducted on the pilot labs and positive feedback has been received.

Data replication is an important technique for improving the accessibility, availability, and reliability of distributed data storage systems. Secret sharing or erasure coding based data fragmentation techniques can be integrated into these systems to improve the system security and survivability. In this paper, a two-tier distributed system is modeled to provide secure data object replication using the integration of data replication and fragmentation approaches. We study the secure replica placement problem in the system with the presence of rational but selfish data owners. Multiple agents place replicas of data objects in the system in order to maximize their own profits while satisfying their security requirements. Algorithmic mechanism design theory is applied to design a direct revelation mechanism to motivate the truth telling of data owners and to approximate the local optimal replica placement. Comparative experiments have been conducted to evaluate the performance of the mechanism.

Innovative learning methodologies and tools are needed to foster students' interest in programming, improve the effectiveness of programming teaching, and retain the students in computing programs. We present a Collaborative Interactive Cyber learning Platform for Anywhere Anytime Java Programming Learning, an open-source cyber learning platform for Computer Science programming education, to support anywhere anytime personalized and collaborative programming learning. A preliminary evaluation of the proposed cyber learning platform and learning methods has been conducted and the students' responses have been very positive.

Cloud computing has tremendous potential to meet the current needs of cyber-education. This paper presents a work-in-progress work to promote personalized and collaborative learning in computer science introductory programming courses through the design and development of a live and interactive cloud learning environment.

This paper proposes to build reusable live programming learning objects for improving introductory computer programming learning. The learning objects employ an innovative learning model, called Interactive Learning Model, to provide visualized and interactive programming learning in a perception and practice integrated, cloud-based, learning platform. A prototype learning environment has been implemented and deployed in Google Cloud to test and evaluate the learning outcomes of proposed learning objects and learning model.