Qiuhao Zeng

Qiuhao Zeng, Wei Wang, Fan Zhou et al.

Existing domain generalization aims to learn a generalizable model to perform well even on unseen domains. For many real-world machine learning applications, the data distribution often shifts gradually along domain indices. For example, a self-driving car with a vision system drives from dawn to dusk, with the sky darkening gradually. Therefore, the system must be able to adapt to changes in ambient illumination and continue to drive safely on the road. In this paper, we formulate such problems as Evolving Domain Generalization, where a model aims to generalize well on a target domain by discovering and leveraging the evolving pattern of the environment. We then propose Directional Domain Augmentation (DDA), which simulates the unseen target features by mapping source data as augmentations through a domain transformer. Specifically, we formulate DDA as a bi-level optimization problem and solve it through a novel meta-learning approach in the representation space. We evaluate the proposed method on both synthetic datasets and realworld datasets, and empirical results show that our approach can outperform other existing methods.

Jinhao Zhang, Kangfei Zhao, Qiuhao Zeng et al.

Transformer-based architectures have become the dominant paradigm for Continuous-Time Dynamic Graph (CTDG) learning, yet their performance remains limited on temporally shifted datasets. In this work, we identify attention dispersion as a shared failure mode of dynamic graph Transformers under temporal distribution shift. Through controlled ablation contrasting structurally and temporally distinguished historical neighbors against random ones, we show that prediction depends on a class of critical nodes that carry consistently more predictive signal than arbitrary neighbors. However, existing Transformers fail to focus on these nodes even when they are present in the input, as temporal shift weakens attention contrast and produces overly dispersed attention distributions. This diagnosis suggests a simple and transferable fix: replace standard attention with differential attention, which suppresses common-mode attention and amplifies distinctive token-level signals. When added to three representative CTDG Transformer baselines, differential attention consistently improves performance, with gains concentrated on high-shift datasets. Attention-level measurements further confirm the mechanism, showing reduced attention entropy and increased attention mass on critical nodes. Building on these findings, we introduce DiffDyG, a reference implementation combining differential attention with standard input encodings. Across 9 benchmarks and three negative sampling protocols, DiffDyG achieves SOTA performance, with especially large gains on the most shifted datasets.

Jerry Huang, Peng Lu, Qiuhao Zeng

Recent advances in natural language processing (NLP) have opened up greater opportunities to enable fine-tuned large language models (LLMs) to behave as more powerful interactive agents through improved instruction-following ability. However, understanding how this impacts confidence calibration for reliable model output has not been researched in full. In this work, we examine various open-sourced LLMs, identifying significant calibration degradation after instruction tuning in each. Seeking a practical solution, we look towards label smoothing, which has been shown as an effective method to regularize for overconfident predictions but has yet to be widely adopted in the supervised fine-tuning (SFT) of LLMs. We first provide insight as to why label smoothing is sufficient to maintain calibration throughout the SFT process. However, settings remain where the effectiveness of smoothing is severely diminished, in particular the case of large vocabulary LLMs (LV-LLMs). We posit the cause to stem from the ability to become over-confident, which has a direct relationship with the hidden size and vocabulary size, and justify this theoretically and experimentally. Finally, we address an outstanding issue regarding the memory footprint of the cross-entropy loss computation in the label smoothed loss setting, designing a customized kernel to dramatically reduce memory consumption without sacrificing speed or performance in comparison to existing solutions for non-smoothed losses.

Zihao Jing, Qiuhao Zeng, Ruiyi Fang et al.

Large language models (LLMs) are enabling reasoning over biomolecular structures, yet existing methods remain modality-specific and typically compress structural inputs through sequence-based tokenization or fixed-length query connectors. Such architectures either omit the geometric groundings requisite for mitigating structural hallucinations or impose inflexible modality fusion bottlenecks that concurrently over-compress and suboptimally allocate structural tokens, thereby impeding the realization of generalized all-atom reasoning. We introduce Cuttlefish, a unified all-atom LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, Scaling-Aware Patching leverages an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, Geometry Grounding Adapter refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across diverse all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning. Code is available at the project repository.

Yi Ding, Neethu Robinson, Chengxuan Tong et al.

Neuropsychological studies suggest that co-operative activities among different brain functional areas drive high-level cognitive processes. To learn the brain activities within and among different functional areas of the brain, we propose LGGNet, a novel neurologically inspired graph neural network, to learn local-global-graph representations of electroencephalography (EEG) for Brain-Computer Interface (BCI). The input layer of LGGNet comprises a series of temporal convolutions with multi-scale 1D convolutional kernels and kernel-level attentive fusion. It captures temporal dynamics of EEG which then serves as input to the proposed local and global graph-filtering layers. Using a defined neurophysiologically meaningful set of local and global graphs, LGGNet models the complex relations within and among functional areas of the brain. Under the robust nested cross-validation settings, the proposed method is evaluated on three publicly available datasets for four types of cognitive classification tasks, namely, the attention, fatigue, emotion, and preference classification tasks. LGGNet is compared with state-of-the-art methods, such as DeepConvNet, EEGNet, R2G-STNN, TSception, RGNN, AMCNN-DGCN, HRNN and GraphNet. The results show that LGGNet outperforms these methods, and the improvements are statistically significant (p<0.05) in most cases. The results show that bringing neuroscience prior knowledge into neural network design yields an improvement of classification performance. The source code can be found at https://github.com/yi-ding-cs/LGG

Yi Ding, Neethu Robinson, Su Zhang et al.

The high temporal resolution and the asymmetric spatial activations are essential attributes of electroencephalogram (EEG) underlying emotional processes in the brain. To learn the temporal dynamics and spatial asymmetry of EEG towards accurate and generalized emotion recognition, we propose TSception, a multi-scale convolutional neural network that can classify emotions from EEG. TSception consists of dynamic temporal, asymmetric spatial, and high-level fusion layers, which learn discriminative representations in the time and channel dimensions simultaneously. The dynamic temporal layer consists of multi-scale 1D convolutional kernels whose lengths are related to the sampling rate of EEG, which learns the dynamic temporal and frequency representations of EEG. The asymmetric spatial layer takes advantage of the asymmetric EEG patterns for emotion, learning the discriminative global and hemisphere representations. The learned spatial representations will be fused by a high-level fusion layer. Using more generalized cross-validation settings, the proposed method is evaluated on two publicly available datasets DEAP and MAHNOB-HCI. The performance of the proposed network is compared with prior reported methods such as SVM, KNN, FBFgMDM, FBTSC, Unsupervised learning, DeepConvNet, ShallowConvNet, and EEGNet. TSception achieves higher classification accuracies and F1 scores than other methods in most of the experiments. The codes are available at https://github.com/yi-ding-cs/TSception

Yi Ding, Neethu Robinson, Qiuhao Zeng et al.

In this paper, we propose a deep learning framework, TSception, for emotion detection from electroencephalogram (EEG). TSception consists of temporal and spatial convolutional layers, which learn discriminative representations in the time and channel domains simultaneously. The temporal learner consists of multi-scale 1D convolutional kernels whose lengths are related to the sampling rate of the EEG signal, which learns multiple temporal and frequency representations. The spatial learner takes advantage of the asymmetry property of emotion responses at the frontal brain area to learn the discriminative representations from the left and right hemispheres of the brain. In our study, a system is designed to study the emotional arousal in an immersive virtual reality (VR) environment. EEG data were collected from 18 healthy subjects using this system to evaluate the performance of the proposed deep learning network for the classification of low and high emotional arousal states. The proposed method is compared with SVM, EEGNet, and LSTM. TSception achieves a high classification accuracy of 86.03%, which outperforms the prior methods significantly (p<0.05). The code is available at https://github.com/deepBrains/TSception

Qiuhao Zeng, Wei Wang, Fan Zhou et al.

In the field of domain generalization, the task of constructing a predictive model capable of generalizing to a target domain without access to target data remains challenging. This problem becomes further complicated when considering evolving dynamics between domains. While various approaches have been proposed to address this issue, a comprehensive understanding of the underlying generalization theory is still lacking. In this study, we contribute novel theoretic results that aligning conditional distribution leads to the reduction of generalization bounds. Our analysis serves as a key motivation for solving the Temporal Domain Generalization (TDG) problem through the application of Koopman Neural Operators, resulting in Temporal Koopman Networks (TKNets). By employing Koopman Operators, we effectively address the time-evolving distributions encountered in TDG using the principles of Koopman theory, where measurement functions are sought to establish linear transition relations between evolving domains. Through empirical evaluations conducted on synthetic and real-world datasets, we validate the effectiveness of our proposed approach.

Qiuhao Zeng, Jerry Huang, Peng Lu et al.

Over recent years, the Transformer has become a fundamental building block for sequence modeling architectures. Yet at its core is the use of self-attention, whose memory and computational cost grow quadratically with the sequence length $N$, rendering it prohibitively expensive for long sequences. A promising approach is top-$k$ attention, which selects only the $k$ most relevant tokens and achieves performance comparable to vanilla self-attention while significantly reducing space and computational demands. However, causal masks require the current query token to only attend to past tokens, preventing the existing top-$k$ attention method from efficiently searching for the most relevant tokens in parallel, thereby limiting training efficiency. In this work, we propose ZETA, leveraging \textbf{Z}-Order Curves for \textbf{E}fficient \textbf{T}op-$k$ \textbf{A}ttention, to enable parallel querying of past tokens for entire sequences. % in both space and time complexity of $\mathcal{O}(N \log N)$. We first theoretically show that the choice of key and query dimensions involves a trade-off between the curse of dimensionality and the preservation of relative distances after projection. In light of this insight, we propose reducing the dimensionality of keys and queries in contrast to values and further leverage $Z$-order curves to map low-dimensional keys and queries into \emph{one}-dimensional space, which permits parallel sorting, thereby largely improving the efficiency for top-$k$ token selection. Experimental results demonstrate that ZETA matches the performance of standard attention on the synthetic \textsc{Multi-Query Associative Recall} task and outperforms attention and its variants on \textsc{Long Range Arena} and \textsc{WikiText-103} language modeling.

Ruiyi Fang, Bingheng Li, Zhao Kang et al.

Graph Domain Adaptation (GDA) addresses a pressing challenge in cross-network learning, particularly pertinent due to the absence of labeled data in real-world graph datasets. Recent studies attempted to learn domain invariant representations by eliminating structural shifts between graphs. In this work, we show that existing methodologies have overlooked the significance of the graph node attribute, a pivotal factor for graph domain alignment. Specifically, we first reveal the impact of node attributes for GDA by theoretically proving that in addition to the graph structural divergence between the domains, the node attribute discrepancy also plays a critical role in GDA. Moreover, we also empirically show that the attribute shift is more substantial than the topology shift, which further underscores the importance of node attribute alignment in GDA. Inspired by this finding, a novel cross-channel module is developed to fuse and align both views between the source and target graphs for GDA. Experimental results on a variety of benchmarks verify the effectiveness of our method.

Jerry Huang, Peng Lu, Qiuhao Zeng et al.

Ensuring that deep learning models are well-calibrated in terms of their predictive uncertainty is essential in maintaining their trustworthiness and reliability, yet despite increasing advances in foundation model research, the relationship between such large language models (LLMs) and their calibration remains an open area of research. In this work, we look at a critical gap in the calibration of LLMs within multilingual settings, in an attempt to better understand how the data scarcity can potentially lead to different calibration effects and how commonly used techniques can apply in these settings. Our analysis on two multilingual benchmarks, over 29 and 42 languages respectively, reveals that even in low-resource languages, model confidence can increase significantly after instruction-tuning on high-resource language SFT datasets. However, improvements in accuracy are marginal or non-existent, resulting in mis-calibration, highlighting a critical shortcoming of standard SFT for multilingual languages. Furthermore, we observe that the use of label smoothing to be a reasonable method alleviate this concern, again without any need for low-resource SFT data, maintaining better calibration across all languages. Overall, this highlights the importance of multilingual considerations for both training and tuning LLMs in order to improve their reliability and fairness in downstream use.

Zihao Jing, Qiuhao Zeng, Ruiyi Fang et al.

Molecular understanding is central to advancing areas such as scientific discovery, yet Large Language Models (LLMs) struggle to understand molecular graphs effectively. Existing graph-LLM bridges often adapt the Q-Former-style connector with fixed-length static tokens, which is originally designed for vision tasks. These designs overlook stereochemistry and substructural context and typically require costly LLM-backbone fine-tuning, limiting efficiency and generalization. We introduce EDT-Former, an Entropy-guided Dynamic Token Transformer that generates tokens aligned with informative molecular patches, thereby preserving both local and global structural features for molecular graph understanding. Beyond prior approaches, EDT-Former enables alignment between frozen graph encoders and LLMs without tuning the LLM backbone (excluding the embedding layer), resulting in computationally efficient finetuning, and achieves stateof-the-art results on MoleculeQA, Molecule-oriented Mol-Instructions, and property prediction benchmarks (TDC, MoleculeNet), underscoring its effectiveness for scalable and generalizable multimodal molecular understanding

Qiuhao Zeng

In few-shot learning, classifiers are expected to generalize to unseen classes given only a small number of instances of each new class. One of the popular solutions to few-shot learning is metric-based meta-learning. However, it highly depends on the deep metric learned on seen classes, which may overfit to seen classes and fail to generalize well on unseen classes. To improve the generalization, we explore the substructures of classifiers and propose a novel meta-learning algorithm to learn each classifier as a combination of meta-components. Meta-components are learned across meta-learning episodes on seen classes and disentangled by imposing an orthogonal regularizer to promote its diversity and capture various shared substructures among different classifiers. Extensive experiments on few-shot benchmark tasks show superior performances of the proposed method.

Peng Lu, Jerry Huang, Qiuhao Zeng et al.

The quadratic complexity of the attention mechanism in Transformer models has motivated the development of alternative architectures with sub-quadratic scaling, such as state-space models. Among these, Mamba has emerged as a leading architecture, achieving state-of-the-art results across a range of language modeling tasks. However, Mamba's performance significantly deteriorates when applied to contexts longer than those seen during pre-training, revealing a sharp sensitivity to context length extension. Through detailed analysis, we attribute this limitation to the out-of-distribution behaviour of its state-space dynamics, particularly within the parameterization of the state transition matrix $\mathbf{A}$. Unlike recent works which attribute this sensitivity to the vanished accumulation of discretization time steps, $\exp(-\sum_{t=1}^NΔ_t)$, we establish a connection between state convergence behavior as the input length approaches infinity and the spectrum of the transition matrix $\mathbf{A}$, offering a well-founded explanation of its role in length extension. Next, to overcome this challenge, we propose an approach that applies spectrum scaling to pre-trained Mamba models to enable robust long-context generalization by selectively modulating the spectrum of $\mathbf{A}$ matrices in each layer. We show that this can significantly improve performance in settings where simply modulating $Δ_t$ fails, validating our insights and providing avenues for better length generalization of state-space models with structured transition matrices.

Qiuhao Zeng, Tianze Luo, Boyu Wang

Unsupervised domain adaptation (UDA) enables knowledge transfer from the labelled source domain to the unlabeled target domain by reducing the cross-domain discrepancy. However, most of the studies were based on direct adaptation from the source domain to the target domain and have suffered from large domain discrepancies. To overcome this challenge, in this paper, we propose the domain-augmented domain adaptation (DADA) to generate pseudo domains that have smaller discrepancies with the target domain, to enhance the knowledge transfer process by minimizing the discrepancy between the target domain and pseudo domains. Furthermore, we design a pseudo-labeling method for DADA by projecting representations from the target domain to multiple pseudo domains and taking the averaged predictions on the classification from the pseudo domains as the pseudo labels. We conduct extensive experiments with the state-of-the-art domain adaptation methods on four benchmark datasets: Office Home, Office-31, VisDA2017, and Digital datasets. The results demonstrate the superiority of our model.