The burgeoning field of text-to-3D synthesis offers transformative potential in diverse domains such as computer-aided design, gaming, virtual reality, and artistic creation. However, the generation struggles with issues of inconsistency and low resolution, primarily due to the lack of critical visual clues like views and attributes. Furthermore, random constraint in rendering may impair model inference, leading to the Janus problem. In response to these challenges, we introduce HexaDream to produce high-quality 3D content. Hexaview Generation Diffusion Model is designed to merge object types, attributes, and view-specific text into unified latent space. Besides, the feature aggregation attention significantly enhances the detail and consistency of the generated output by mapping point features from orthogonal view into the 3D domain. Another innovation is the Dynamic-weighted HexaConstraint. This module employs a projection matrix to generate projected views and calculates the differential loss between these projections and the hexaviews, ensuring high fidelity. Our comparative experiments show that HexaDream achieves improvements of 8% in CLIP-R, 12% in Keypart Fidelity, and especially 20.6% in Multihead Alleviation compared with existing methods respectively.

Multi-instance multi-label learning is a general framework in which each sample is represented as a bag of instances associated with multiple labels. However, two weaknesses remain and hinder its performance in real-world tasks. One is that bag generators often neglect positional information (e.g., the location of a pixel in the image) when generating bags, making position-related labels indistinguishable. The other is that the MIL assumption does not always hold. In some real-world tasks, labels have hierarchical low-level concepts, and these concepts are related to certain combinations of instances instead of one single instance. In this paper, we propose the Position-Aware Doubly Graph Convolutional Networks (padGCN). On the one hand, padGCN generates bags by arranging instances in a multi-instance graph to aggregate instances’ features by exploiting positional relationships among them. Then instances that aggregate other instances’ features are input into a neural network to obtain sub-sub-concepts used for multi-label learning. On the other hand, padGCN learns sub-concepts from labels and organizes sub-sub-concepts, sub-concepts, and labels in a tripartite multi-label graph in hyperbolic space to exploit their hierarchical structure. Experiments are conducted on 6 image and text data sets. Compared to the SOTA methods, padGCN averagely achieves 5% improvement on 7 measurements. Pair-wise t-test results on 42 experiments indicate that padGCN is significantly better than SOTA methods in 30 experiments, comparable to SOTA methods in 12 experiments, and never worse than SOTA methods, which verifies the superiority and robustness of padGCN. Runtime experiments show that padGCN is comparable to SOTA methods and is computationally efficient.

Utilizing pre-trained generation models for predicting sentiment elements has recently shown significant advancements in aspect sentiment quad prediction benchmarks. However, these models overlook the significance of syntactic information, which have proven to be effective in previous extraction-based approaches. Different from extraction-based models, efficiently encoding the syntactic structure in generation model is challenging because such models are pretrained on natural language, and modeling structured data may lead to catastrophic forgetting of distributional knowledge. In this study, we propose an innovative structure-aware framework that explicitly encodes the syntactic structure into the pre-trained generation model while preserving its original distributional knowledge.

Graphs are structured data that models complex relations between real-world entities. Heterophilic graphs, where linked nodes trend to have different labels or dissimilar features, have recently attracted significant attention and found many real-world applications. Meanwhile, increasing efforts have been made to advance learning from graphs with heterophily. Various graph heterophily measures, benchmark datasets, and learning paradigms are emerging rapidly. In this survey, we comprehensively review existing works on learning from graphs with heterophily. First, we overview over 500 publications, of which more than 300 are directly related to heterophilic graphs. After that, we survey existing metrics of graph heterophily and list recent benchmark datasets. Further, we systematically categorize existing methods based on a hierarchical taxonomy including GNN models, learning paradigms and practical applications. In addition, broader topics related to graph heterophily are also included. Finally, we discuss the primary challenges of existing studies and highlight promising avenues for future research.

The field of time series forecasting has been seen widespread application of Transformer-based architectures. However, the quadratic complexity of the attention mechanism limits its performance in long-term time series forecasting. The proposition of patching mechanism has alleviated this issue to some extent, but models will struggle to effectively unify the information between intra-patch and inter-patch. To address this problem, we propose DualMamba, a novel Mamba-based model for time series forecasting, which segments the time series into subseries-level patches and employs dual Mamba modules to capture local and global information separately. Specifically, the time series use patch-wise dependencies to guide the local module, where each patch uses a point-wise representation of time series data. Furthermore, we design an information fusion mechanism for integrating information between intra-patch and inter-patch, which effectively incorporates global information into local contexts. This allows the model to capture both local details and global trends. Extensive experiments on several real-world datasets demonstrate that DualMamba achieves state-of-the-art performance in most cases and has reliable robustness, making it highly adaptable for various types of time series.

In the domain of quantitative trading, the imperative is to translate historical financial data into predictive signals, commonly referred to as alpha factors, which serves to anticipate future market trends. Notably, formulaic alphas that are expressible via explicit mathematical formulas are highly sought after by certain investors for better interpretability. The evolving landscape of technology has witnessed the increasing deployment of large language models (LLMs) across various domains, which raises the question of whether LLMs can be effective in the context of formulaic alpha-mining tasks. This paper presents several paradigms aimed at integrating LLMs into the optimization loop of alpha mining, including scenarios where an LLM serves as the sole alpha generator, as well as instances where LLMs enhance existing frameworks. Empirical evaluations on real-world stock data demonstrate significant performance improvements, with our hybrid method achieving an average information coefficient (IC) of 0.0515, a 75% improvement over the baseline — a state-of-the-art reinforcement learning-based framework; backtesting further reveals a cumulative excess return more than double the baseline framework. These results underscore the potential of LLM-enhanced approaches in advancing formulaic alpha discovery and driving innovation in quantitative trading.

While deep learning systems demonstrate good performance in many fields such as computer vision, natural language processing, and computational biology, time and data efficiency still remain as two major challenges. Deep multi-task learning, in which one network produces predictive outputs for multiple tasks, has emerged as a promising approach with fast inference and good performance. However, how to balance the learning of each individual task is difficult in deep multi-task learning. In this paper, we present a combinational method called POMSI to project conflicting gradients and mitigate the scale imbalance in multi-task learning. The proposed POMSI method can be trained end-to-end with all kinds of losses without any distributional assumption. Moreover, the POMSI model is model-agnostic and can be applied to existing multi-task architectures for further enhancement. Through extensive experiments on benchmark datasets, the proposed POMSI method achieves substantial gains in the performance compared with state-of-the-art methods.

Machine Unlearning (MU) has emerged as a promising technique for aligning large language models (LLMs) with safety requirements to steer them forgetting specific harmful contents. Despite the significant progress in previous studies, we argue that the current evaluation criteria, which solely focus on safety evaluation, are actually impractical and biased, leading to concerns about the true effectiveness of MU techniques. To address this, we propose to comprehensively evaluate LLMs after MU from three aspects: safety, over-safety, and general utility. Specifically, a novel benchmark MUBENCH with 18 related datasets is first constructed, where the safety is measured with both vanilla harmful inputs and 10 types of jailbreak attacks. Furthermore, we examine whether MU introduces side effects, focusing on over-safety and utility-loss. Extensive experiments are performed on 3 popular LLMs with 7 recent MU methods. The results highlight a challenging trilemma in safety alignment without side effects, indicating that there is still considerable room for further exploration. MUBENCH serves as a comprehensive benchmark, fostering future research on MU for safety alignment of LLMs.

Diversity plays a significant role in many problems, such as ensemble learning, reinforcement learning, and combinatorial optimization. How to define the diversity measure is a longstanding problem. Many methods rely on expert experience to define a proper behavior space and then obtain the diversity measure, which is, however, challenging in many scenarios. In this paper, we propose the problem of learning a behavior space from human feedback and present a general method called Diversity from Human Feedback (DivHF) to solve it. DivHF learns a behavior descriptor consistent with human preference by querying human feedback. The learned behavior descriptor can be combined with any distance measure to define a diversity measure. We demonstrate the effectiveness of DivHF by integrating it with the Quality-Diversity optimization algorithm MAP-Elites and conducting experiments on the QDax suite. The results show that the behavior learned by DivHF is much more consistent with human requirements than the one learned by direct data-driven approaches without human feedback, and makes the final solutions more diverse under human preference. Our contributions include formulating the problem, proposing the DivHF method, and demonstrating its effectiveness through experiments.

In multi-dimensional multi-label classification (MDML), a number of heterogeneous label spaces are assumed to characterize the rich semantics of one object from different dimensions and a set of proper labels can be assigned to the object from each heterogeneous label space. In recent years, similarity-based framework has achieved a promising performance in classification tasks (e.g., multi-class/multi-label classification), while its effectiveness has not been investigated in solving the MDML problems. Moreover, existing similarity-based approaches only utilize either instance-based or label-based information which limits their generalization ability. In this paper, we propose a novel similarity-based MDML approach, naming SIDLE which attempts to utilize both instance-based and label-based information. To extract similarity information, SIDLE first identifies k nearest neighbors in instance space and enhanced label space, respectively. Then, with these identified samples, SIDLE calculates the simple counting statistics based on their labels as well as a bias based on distance between the sample and these identified samples. Finally, the instance space is enriched with extracted similarity information to update instance space and enhanced label space. These three steps are iteratively conducted until convergence. Experiments validate the effectiveness of the proposed SIDLE approach.