Pengfei Li, Shijie Wang, Fangyuan Li, Yikun Fu, Kaifeng Liu, Kaiyan Zhang, Dazhi Zhang, Yuqiang Li, Biqing Qi, Bowen Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Heterogeneous Large Language Model (LLM) fusion integrates the strengths of multiple source LLMs with different architectures into a target LLM with low computational overhead. While promising, existing methods suffer from two major limitations: 1) **reliance on real data from limited domain** for knowledge fusion, preventing the target LLM from fully acquiring knowledge across diverse domains, and 2) **fixed data allocation proportions** across domains, failing to dynamically adjust according to the target LLM's varying capabilities across domains, leading to a capability imbalance. To overcome these limitations, we propose Bohdi, a synthetic-data-only heterogeneous LLM fusion framework. Through the organization of knowledge domains into a hierarchical tree structure, Bohdi enables automatic domain exploration and multi-domain data generation through multi-model collaboration, thereby comprehensively extracting knowledge from source LLMs. By formalizing domain expansion and data sampling proportion allocation on the knowledge tree as a Hierarchical Multi-Armed Bandit problem, Bohdi leverages the designed DynaBranches mechanism to adaptively adjust sampling proportions based on the target LLM's performance feedback across domains. Integrated with our proposed Introspection-Rebirth (IR) mechanism, DynaBranches dynamically tracks capability shifts during target LLM's updates via Sliding Window Binomial Likelihood Ratio Testing (SWBLRT), further enhancing its online adaptation capability. Comparative experimental results on a comprehensive suite of benchmarks demonstrate that Bohdi significantly outperforms existing baselines on multiple target LLMs, exhibits higher data efficiency, and virtually eliminates the imbalance in the target LLM's capabilities.

Pansharpening aims to recover the spectral information and spatial details of high-resolution multispectral (HRMS) images with high accuracy. In this context, we design a triple-scale multi-level network guided by Hessian spatial loss for pansharpening (HTMP). Firstly, we propose a novel Hessian spatial loss designed to establish deep mapping relationships in the spatial domain. Hessian spatial loss guides the network in enhancing its ability to characterize the edges of blurred regions while maintaining the consistency of spatial details. Secondly, we employ a triple-scale multi-level feature extraction network (TMFENet) to obtain comprehensive spectral and spatial features, thereby enhancing the interaction of multi-scale contextual information from coarse-grained to fine-grained scales. To efficiently integrate and represent cross-modal long-distance information, the spectral and spatial features are treated as a whole and input into the triple-scale adaptive sparse transformer (TASTrans) to extract global features. Finally, the multi-level image reconstruction network (MIRNet) combines multi-modal features at different resolutions to progressively generate HRMS images rich in semantics. Experiments performed on datasets demonstrate that our method produces fused images with superior visual quality compared to state-of-the-art methods. Furthermore, it is obvious from the ablation experiments that incorporating the Hessian spatial loss significantly enhances the fusion performance of deep learning models.

In the hyperspectral image classification (HSIC) task, the most commonly used model validation paradigm is partitioning the training-test dataset through pixelwise random sampling. By training on a small amount of data, the deep learning model can achieve almost perfect accuracy. However, in our experiments, we found that the high accuracy was reached because the training and test datasets share a lot of information. On nonoverlapping dataset partitions, well-performing models suffer significant performance degradation. To this end, we propose a spectral-spatial axial aggregation transformer model, namely, SaaFormer, which preserves generalization across dataset partitions. SaaFormer applies a multilevel spectral extraction structure to segment the spectrum into multiple spectrum clips such that the wavelength continuity of the spectrum across the channel is preserved. For each spectrum clip, the axial aggregation attention mechanism, which integrates spatial features along multiple spectral axes, is applied to mine the spectral characteristic. The multilevel spectral extraction and the axial aggregation attention emphasize spectral characteristics to improve the model generalization. The experimental results on five publicly available datasets demonstrate that our model exhibits comparable performance on the random partition while significantly outperforming other methods on nonoverlapping partitions. Moreover, SaaFormer shows excellent performance on background classification.

The transferability of adversarial perturbations provides an effective shortcut for black-box attacks. Targeted perturbations have greater practicality but are more difficult to transfer between models. In this paper, we experimentally and theoretically demonstrated that neural networks trained on the same dataset have more consistent performance in High-Sample-Density-Regions (HSDR) of each class instead of low sample density regions. Therefore, in the target setting, adding perturbations towards HSDR of the target class is more effective in improving transferability. However, density estimation is challenging in high-dimensional scenarios. Further theoretical and experimental verification demonstrates that easy samples with low loss are more likely to be located in HSDR. Perturbations towards such easy samples in the target class can avoid density estimation for HSDR location. Based on the above facts, we verified that adding perturbations to easy samples in the target class improves targeted adversarial transferability of existing attack methods. A generative targeted attack strategy named Easy Sample Matching Attack (ESMA) is proposed, which has a higher success rate for targeted attacks and outperforms the SOTA generative method. Moreover, ESMA requires only $5\%$ of the storage space and much less computation time comparing to the current SOTA, as ESMA attacks all classes with only one model instead of seperate models for each class. Our code is available at https://github.com/gjq100/ESMA

Image deformation refers to deforming objects in images into a target shape or posture. Although point-based image deformation algorithms have made breakthroughs in performance and visual effects, they are limited to warping the current structure information in the source image. For example, while opening a closed mouth in a face deformation, the point-based algorithm cannot generate teeth, which causes the mouth to twist weirdly. Deep learning-based face editing models can generate new parts but cannot achieve fine pixel-level manipulation. To overcome these challenges, we propose a two-step strategy for face deformation. We first generate a high-resolution intermediate image by blending the source image and the specific part of the target image via our face blending generative adversarial network (FB-GAN). Then, we employ a state-of-the-art point-based geometric deformation method to deform the intermediate image with target face guidance. Extensive experiments show that the proposed FB-GAN can generate realistic and high-resolution results and demonstrates that the two-step face deformation strategy can be better applied to human face deformation.

This paper considers colorization-based image compression in RGB color space. In compression, we store only the compressed luminance component of the original color image and a few representative pixels extracted from the original color image. In decompression, by explicitly introducing the relation between the luminance component and the original color image into diffusion equations, a linear reaction-diffusion system with Perona--Malik type diffusion coefficient is proposed to reconstruct R, G, and B channels simultaneously. The Perona--Malik type diffusion coefficient is a function of the luminance component and leads to interior degenerations, in general. It yields anisotropic smoothing in the restored color image and constrains the geometry of the restored image to follow the geometry of the luminance component. The existence and uniqueness of solutions for the proposed system with a specific class of diffusion coefficients are proved in a weighted Sobolev space. The selection of representative pixels has a big impact on reconstruction results. We also propose a local-optimal strategy that splits the original color image into a series of different size subimages and searches the optimal representative pixel in each subimage. Comparisons with recent colorization-based image compression methods, as well as transform-based JPEG and JPEG2000 standards, are performed to show the potential for successful compression applications of the proposed method.

Multiplicative noise removal is a challenging task in image processing. Inspired by the impressive performance of nonlinear diffusion models in additive noise removal, we address this problem in the view of nonlinear diffusion equation theories rather than the traditional variation methods. We develop a nonlinear diffusion filter denoising framework, which considers not only the information of the gradient of the image, but also the information of gray levels of the image. Furthermore, under this framework, we propose a doubly degenerate diffusion model for multiplicative noise removal, which is analyzed with respect to some of its properties and behavior in denoising process. In numerical aspects, we present an efficient scheme which uses a stabilization by fast explicit diffusion for the implementation of the multiplicative noise removal model. Finally, the experimental results illustrate effectiveness and efficiency of the proposed model.

Complex Event Processing (CEP) over event streams has become increasingly important for real-time applications ranging from healthcare to supply chain management. In such applications, arbitrarily complex sequence patterns as well as non existence of such complex situations must be detected in real time. To assure real-time responsiveness for detection of such complex pattern over high volume high-speed streams, efficient processing techniques must be designed. Unfortunately the efficient processing of complex sequence queries with negations remains a largely open problem to date. To tackle this shortcoming, we designed optimized strategies for handling nested CEP query. In this demonstration, we propose to showcase these techniques for processing and optimizing nested pattern queries on streams. In particular our demonstration showcases a platform for specifying complex nested queries, and selecting one of the alternative optimized techniques including sub-expression sharing and intermediate result caching to process them. We demonstrate the efficiency of our optimized strategies by graphically comparing the execution time of the optimized solution against that of the default processing strategy of nested CEP queries. We also demonstrate the usage of the proposed technology in several healthcare services.

This paper introduces a class of adaptive Perona-Malik (PM) diffusion, which combines the PM equation with the heat equation. The PM equation provides a potential algorithm for image segmentation, noise removal, edge detection, and image enhancement. However, the defect of traditional PM model is tending to cause the staircase effect and create new features in the processed image. Utilizing the edge indicator as a variable exponent, we can adaptively control the diffusion mode, which alternates between PM diffusion and Gaussian smoothing in accordance with the image feature. Computer experiments indicate that the present algorithm is very efficient for edge detection and noise removal.

We propose a method for vector field learning with outliers, called vector field consensus (VFC). It could distinguish inliers from outliers and learn a vector field fitting for the inliers simultaneously. A prior is taken to force the smoothness of the field, which is based on the Tiknonov regularization in vector-valued reproducing kernel Hilbert space. Under a Bayesian framework, we associate each sample with a latent variable which indicates whether it is an inlier, and then formulate the problem as maximum a posteriori problem and use Expectation Maximization algorithm to solve it. The proposed method possesses two characteristics: 1) robust to outliers, and being able to tolerate 90% outliers and even more, 2) computationally efficient. As an application, we apply VFC to solve the problem of mismatch removing. The results demonstrate that our method outperforms many state-of-the-art methods, and it is very robust.

Buffer overflow vulnerabilities are program defects that can cause a buffer to overflow at runtime. Many security attacks exploit buffer overflow vulnerabilities to compromise critical data structures. In this paper, we present a black-box testing approach to detecting buffer overflow vulnerabilities. Our approach is motivated by a reflection on how buffer overflow vulnerabilities are exploited in practice. In most cases the attacker can influence the behavior of a target system only by controlling its external parameters. Therefore, launching a successful attack often amounts to a clever way of tweaking the values of external parameters. We simulate the process performed by the attacker, but in a more systematic manner. A novel aspect of our approach is that it adapts a general software testing technique called combinatorial testing to the domain of security testing. In particular, our approach exploits the fact that combinatorial testing often achieves a high level of code coverage. We have implemented our approach in a prototype tool called Tance. The results of applying Tance to five open-source programs show that our approach can be very effective in detecting buffer overflow vulnerabilities.

In this paper, we present a robust and efficient approach to extract motion layers from a pair of images with large disparity motion. First, motion models are established as: 1) initial SIFT matches are obtained and grouped into a set of clusters using our developed topological clustering algorithm; 2) for each cluster with no less than three matches, an affine transformation is estimated with least-square solution as tentative motion model; and 3) the tentative motion models are refined and the invalid models are pruned. Then, with the obtained motion models, a graph cuts based layer assignment algorithm is employed to segment the scene into several motion layers. Experimental results demonstrate that our method can successfully segment scenes containing objects with large interframe motion or even with significant interframe scale and pose changes. Furthermore, compared with the previous method invented by Wills and its modified version, our method is much faster and more robust.

Attestation has become a promising approach for ensuring software integrity in wireless sensor networks. However, current attestation either focuses on static system properties, e.g., code integrity, or requires hardware support such as Trusted Platform Module (TPM). However, there are attacks exploiting vulnerabilities that do not violate static system properties, and sensor platforms may not have hardware-based security support. This paper presents a software attestation scheme for dynamic data integrity based on data boundary integrity. It automatically transforms the source code and inserts data guards to track run-time program data. A data guard is unrecoverable once it is corrupted by an attacker, even if the attacker fully controls the system later. The corruption of any data guard at runtime can be remotely detected. A corruption either indicates a software attack or a bug in the software that needs immediate attention. The benefits of the proposed attestation scheme are as follows. First, it does not rely on any additional hardware support, making it suitable for low-cost sensor nodes. Second, it introduces minimal communication cost and has adjustable runtime memory overhead. Third, it works even if sensor nodes use different hardware platforms, as long as they run the same software. The prototype implementation and the experiments on TelosB motes show that the proposed technique is both effective and efficient for sensor networks.

Security testing has gained significant attention recently due to frequent attacks against software systems. This paper presents a trace-based security testing approach. It reuses test cases generated from previous testing methods to produce execution traces. An execution trace is a sequence of program statements exercised by a test case. Each trace is symbolically executed to produce program constraints and security constraints. A program constraint is a constraint imposed by program logic on program variables. A security constraint is a condition on program variables that must be satisfied to ensure system security. A security flaw exists if there is an assignment of values to program variables that satisfies the program constraint but violates the security constraint. This approach detects security flaws even if existing test cases do not trigger them. The novelty of this method is a test model that unifies program constraints and security constraints such that formal reasoning can be applied to detect vulnerabilities. A tool named SecTAC is implemented and applied to 14 benchmark programs and 3 open-source programs. The experiment shows that SecTAC quickly detects all reported vulnerabilities and 13 new ones that have not been detected before.

Traditional test case generation methods usually consider coverage criteria like statement or path coverage and ignore security characteristics. The result is that a test case may fail to find vulnerabilities even if it covers the vulnerable statements. However, we argue that existing test cases are still of great value because significant human effort and time have been invested to achieve high coverage criteria. A high coverage indicates a high possibility that vulnerable statements occur in the execution traces of these test cases. Thus existing test cases could guide us to those vulnerable statements. Under this intuition, we present a method of security testing by re-examining existing test cases. The basic idea is to discover two types of constraints in a program: program constraints (PC) and security constraints (SC). The former are the constraints imposed by program statements. For example, an assignment statement i=0 constrains the value of i to be 0. The later are the constraints derived from security concerns. For example, a buffer should never be overflowed. Intuitively, a statement is vulnerable if it can make PCrarrSC be false, which means the program constraints are not strict enough to ensure the security constraints. We design and develop a tool named RETAST to demonstrate our idea and the initial result is promising.

In this article, an inference process is defined as a series of events in which the truth values flow from propositions along certain inference channels. the concepts of netlike inference process and solution searching process are then described. the notion of excitedness is defined as a measure of the activeness of thinking. In the context of an inference process, excitedness describes the truth of the proposition or the belief in the proposition. While in a solution searching process, excitedness describes the ability and/or desire to solve the problem. By introducing simple flows and their network graphs, the process of excitedness flows on the network is described be a set of differential equations with steady state solutions and stability analysis performed by applying Markov process theory. By introducing the concepts of complex flows and multi-branch graphs, the process of excitedness flows on the graph is also described by a set of differential equations with steady state solutions and stability analysis performed similar to Prigogine's theory of dissipative structures.1 Finally, the idea of using computers in netlike inference is proposed.