Recent advances in large language models (LLMs) have intensified efforts to fuse heterogeneous open-source models into a unified system that inherits their complementary strengths. Existing logit-based fusion methods maintain inference efficiency but treat vocabulary dimensions independently, overlooking semantic dependencies encoded by cross-dimension interactions. These dependencies reflect how token types interact under a model's internal reasoning and are essential for aligning models with diverse generation behaviors. To explicitly model these dependencies, we propose InfiGFusion, the first structure-aware fusion framework with a novel Graph-on-Logits Distillation (GLD) loss. Specifically, we retain the top-k logits per output and aggregate their outer products across sequence positions to form a global co-activation graph, where nodes represent vocabulary channels and edges quantify their joint activations. To ensure scalability and efficiency, we design a sorting-based closed-form approximation that reduces the original O(n4)cost of Gromov-Wasserstein distance to O(nlogn), with provable approximation guarantees. Experiments across multiple fusion settings show that GLD consistently improves fusion quality and stability. InfiGFusion outperforms SOTA models and fusion baselines across 11 benchmarks spanning reasoning, coding, and mathematics. It shows particular strength in complex reasoning tasks, with +35.6 improvement on Multistep Arithmetic and +37.06 on Causal Judgement over SFT, demonstrating superior multi-step and relational inference.

Adversarial attacks are widely used to evaluate model robustness, and in black-box scenarios, the transferability of these attacks becomes crucial. Existing generatorbased attacks have excellent generalization and transferability due to their instanceagnostic nature. However, when training generators for multi-target tasks, the success rate of transfer attacks is relatively low due to the limitations of the model's capacity. To address these challenges, we propose a novel Dual-Flow framework for multi-target instance-agnostic adversarial attacks, utilizing Cascading Distribution Shift Training to develop an adversarial velocity function. Extensive experiments demonstrate that Dual-Flow significantly improves transferability over previous multi-target generative attacks. For example, it increases the success rate from Inception-v3 to ResNet-152 by 34.58%. Furthermore, our attack method shows substantially stronger robustness against defense mechanisms, such as adversarially trained models. The code of Dual-Flow is available at: https://github.com/Chyxx/Dual-Flow.

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 multimodel 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. Our code is available at Bohdi.

Model fusion combines multiple Large Language Models (LLMs) with different strengths into a more powerful, integrated model through lightweight training methods.

In this paper, we investigate source-free domain adaptation (SFDA) for visible-infrared person re-identification (VI-ReID), aiming to adapt a pre-trained source model to an unlabeled target domain without access to source data. To address this challenging setting, we propose a novel learning paradigm, termed Source-Free Visible-Infrared Person Re-Identification (SVIP), which fully exploits the prior knowledge embedded in the source model to guide target domain adaptation. The proposed framework comprises three key components specifically designed for the source-free scenario: 1) a Source-Guided Contrastive Learning (SGCL) module, which leverages the discriminative feature space of the frozen source model as a reference to perform contrastive learning on the unlabeled target data, thereby preserving discrimination without requiring source samples; 2) a Residual Transfer Learning (RTL) module, which learns residual mappings to adapt the target model's representations while maintaining the knowledge from the source model; and 3) a Structural Consistency-Guided Cross-modal Alignment (SCCA) module, which enforces reciprocal structural constraints between visible and infrared modalities to identify reliable cross-modal pairs and achieve robust modality alignment without source supervision. Extensive experiments on benchmark datasets demonstrate that SVIP substantially enhances target domain performance and outperforms existing unsupervised VI-ReID methods under source-free settings.

Model fusion combines multiple Large Language Models (LLMs) with different strengths into a more powerful, integrated model through lightweight training methods.

Empirical results in Section 3.1 in the main paper show that simply removing ReLUs lead to improved transferability. In this section, we try freezing all learnable parameters in the unmodified sub-net h during fine-tuning and a similar observation about the initial improvement of transferability can still be decrease made and (see finally Figure the 5). Classification loss of these modified VGG-19 models on the benign CIFAR-10 test set is also reported, in Figure 6. On ImageNet, it is evaluated on the 50000official validation images. As mentioned in the main paper, many recent successes in improving adversarial transferability benefit from maximizing intermediate level distortions rather than the final prediction losses [8, 3, 2] of DNNs.

The vulnerability of deep neural networks (DNNs) to adversarial examples has drawn great attention from the community. In this paper, we study the transferability of such examples, which lays the foundation of many black-box attacks on DNNs. We revisit a not so new but definitely noteworthy hypothesis of Goodfellow et al.'s and disclose that the transferability can be enhanced by improving the linearity of DNNs in an appropriate manner. We introduce linear backpropagation (LinBP), a method that performs backpropagation in a more linear fashion using off-the-shelf attacks that exploit gradients. More specifically, it calculates forward as normal but backpropagates loss as if some nonlinear activations are not encountered in the forward pass. Experimental results demonstrate that this simple yet effective method obviously outperforms current state-of-the-arts in crafting transferable adversarial examples on CIFAR-10 and ImageNet, leading to more effective attacks on a variety of DNNs.

Pre-trained language models have been successful on text classification tasks, but are prone to learning spurious correlations from biased datasets, and are thus vulnerable when making inferences in a new domain. Prior work reveals such spurious patterns via post-hoc explanation algorithms which compute the importance of input features. Further, the model is regularized to align the importance scores with human knowledge, so that the unintended model behaviors are eliminated. However, such a regularization technique lacks flexibility and coverage, since only importance scores towards a pre-defined list of features are adjusted, while more complex human knowledge such as feature interaction and pattern generalization can hardly be incorporated. In this work, we propose to refine a learned language model for a target domain by collecting human-provided compositional explanations regarding observed biases. By parsing these explanations into executable logic rules, the human-specified refinement advice from a small set of explanations can be generalized to more training examples. We additionally introduce a regularization term allowing adjustments for both importance and interaction of features to better rectify model behavior. We demonstrate the effectiveness of the proposed approach on two text classification tasks by showing improved performance in target domain as well as improved model fairness after refinement1.

Pc i=1 yi = 1is satisfied, otherwise f (y) = by duality. A.2 Experiments on Binary Classification with Exponential Loss Here we present the results on a binary classification task over a synthetic dataset of 100 dimensional gaussian clusters. For Σ, similar to [23], we sample a diagonal matrix D, where each entry is sampled uniformly from a specified range, and a rotation matrix U from a HAAR distribution, giving Σ = UDUT. For the source data, we sample µ 1s,µ+1s,Σ 1s,Σ+1sas specified above with k = 0. Now to create a distribution shifted data of various severity, we sample µ 1t,µ+1t,Σ 1t,Σ+1tas specified above with k = 1, which are then used to sample the shifted data as follows: Exponential Loss for Binary Classification Let z be the classification score hθ(x). For logistic training loss, conjugate adaptation loss would default to entropy with sigmoid probability.