This paper focuses on the newly emerged research topic, i.e., whether the complex decision-making logic of a DNN can be mathematically summarized into a few simple logics. Beyond the explanation of a static DNN, in this paper, we hope to show that the seemingly complex learning dynamics of a DNN can be faithfully represented as the change of a few primitive interaction patterns encoded by the DNN. Therefore, we redefine the interaction of principal feature components in intermediate-layer features, which enables us to concisely summarize the highly complex dynamics of interactions throughout the learning of the DNN. The mathematical faithfulness of the new interaction is experimentally verified. From the perspective of learning efficiency, we find that the interactions naturally belong to five groups (reliable, withdrawn, forgotten, betraying, and fluctuating interactions), each representing a distinct type of dynamics of an interaction being learned and/or being forgotten. This provides deep insights into the learning process of a DNN.

As a central component in formal verification, statement autoformalization has been widely studied including the recent efforts from machine learning community, but still remains a widely-recognized difficult and open problem. In this paper, we delve into two critical yet under-explored gaps: 1) absence of faithful and universal automated evaluation for autoformalization results; 2) agnosia of contextual information, inducing severe hallucination of formal definitions and theorems. To address the first issue, we propose **BEq** (_**B**idirectional **E**xtended Definitional E**q**uivalence_), an automated neuro-symbolic method to determine the equivalence between two formal statements, which is formal-grounded and well-aligned with human intuition. For the second, we propose **RAutoformalizer** (_**R**etrieval-augmented **Autoformalizer**_), augmenting statement autoformalization by _Dependency Retrieval_, retrieving potentially dependent objects from formal libraries. We parse the dependencies of libraries and propose to _structurally informalise_ formal objects by the topological order of dependencies. To evaluate OOD generalization and research-level capabilities, we build a novel benchmark, _Con-NF_, consisting of 961 informal-formal statement pairs from frontier mathematical researches. Experiments validate the effectiveness of our approaches: BEq is evaluated on 200 diverse formal statement pairs with expert-annotated equivalence label, exhibiting significantly improved accuracy ($82.50\\% \mapsto 90.50\\%$) and precision ($70.59\\% \mapsto 100.0\\%$). For dependency retrieval, a strong baseline is devised. Our RAutoformalizer substantially outperforms SOTA baselines in both in-distribution ProofNet benchmark ($12.83\\% \mapsto 18.18\\%$, BEq@8) and OOD Con-NF scenario ($4.58\\%\mapsto 16.86\\%$, BEq@8).

This paper focuses on the fundamental challenge of partitioning input variables in attribution methods for Explainable AI, particularly in Shapley value-based approaches. Previous methods always compute attributions given a predefined partition but lack theoretical guidance on how to form meaningful variable partitions. We identify that attribution conflicts arise when the attribution of a coalition differs from the sum of its individual variables' attributions. To address this, we analyze the numerical effects of AND-OR interactions in AI models and extend the Shapley value to a new attribution metric for variable coalitions. Our theoretical findings reveal that specific interactions cause attribution conflicts, and we propose three metrics to evaluate coalition faithfulness. Experiments on synthetic data, NLP, image classification, and the game of Go validate our approach, demonstrating consistency with human intuition and practical applicability.

Hierarchical text classification (HTC) is an important yet challenging task in natural language processing (NLP), primarily due to the complexity of its taxonomic label hierarchy. Existing methods often utilize graph neural networks (GNNs) to model the hierarchical label structure, with the aim of integrating hierarchical information to improve model performance. However, these methods typically fail to fully exploit the topological features of hierarchical labels and are largely dependent on prior knowledge, which limits the generalization ability of the model. To address this limitation, we explore a model with strong generalization ability that improves HTC performance by leveraging topological features, without relying on prior statistics or label semantics. In this paper, we formalize the hierarchical structure as a disconnected graph and introduce a Persistent Homology-based Graph Attention Network (PHGAT) to capture the multi-scale topological features of hierarchical label structure. Building upon this, we propose HiGATA, which exclusively utilizes the topological information of the label hierarchy. By integrating persistent homology-based topological features into node representations, HiGATA further improves text classification performance. We validate the effectiveness of HiGATA through experiments on three widely used datasets.

Deductive formal problem-solving (D-FPS) enables process-verified, human-aligned problem-solving by implementing deductive solving processes within formal theorem proving (FTP) environments. However, current methods fail to address the misalignment between informal and formal reasoning granularity and suffer from inefficiency due to backtracking and error propagation. Moreover, the extreme scarcity of formal problem-solution pairs further hinders progress. For the first gap, we propose **HAR** (_**H**ierarchical **A**utoregressive Formal **R**easoner_), a novel reasoning pipeline. HAR decouples informal-aligned drafting and detailed proving, and formulates solution construction as autoregressive generation with per-step feedback. Second, we propose **CoPA** (_**C**hain-**o**f-**P**roxy-**A**utoformalization_), a data generation pipeline that cascades statement autoformalization, proof drafting, and proof search as a proxy autoformalization path. Experiments demonstrate significant improvements: trained on data bootstrapped by CoPA, HAR achieves superior performance on FormalMath500 ($15.50\\%\mapsto 44.09\\%$) and MiniF2F-Solving ($21.87\\%\mapsto 56.58\\%$) with lower computational budget. Explorations reveal promising directions in formal solution pruning and informal dataset denoising.

The Boolean Satisfiability Problem (SAT) plays a crucial role in cryptanalysis, enabling tasks like key recovery and distinguisher construction. Conflict-Driven Clause Learning (CDCL) has emerged as the dominant paradigm in modern SAT solving, and machine learning has been increasingly integrated with CDCL-based SAT solvers to tackle complex cryptographic problems. However, the lack of a unified evaluation framework, inconsistent input formats, and varying modeling approaches hinder fair comparison. Besides, cryptographic SAT instances also differ structurally from standard SAT problems, and the absence of standardized datasets further complicates evaluation. To address these issues, we introduce SAT4CryptoBench, the first comprehensive benchmark for assessing machine learning–based solvers in cryptanalysis. SAT4CryptoBench provides diverse SAT datasets in both Arithmetic Normal Form (ANF) and Conjunctive Normal Form (CNF), spanning various algorithms, rounds, and key sizes. Our framework evaluates three levels of machine learning integration: standalone distinguishers for instance classification, heuristic enhancement for guiding solving strategies, and hyperparameter optimization for adapting to specific problem distributions. Experiments demonstrate that ANF-based networks consistently achieve superior performance over CNF-based networks in learning cryptographic features. Nonetheless, current ML techniques struggle to generalize across algorithms and instance sizes, with computational overhead potentially offsetting benefits on simpler cases. Despite this, ML-driven optimization strategies notably improve solver efficiency on cryptographic SAT instances. Besides, we propose BASIN, a bitwise solver taking plaintext-ciphertext bitstrings as inputs. Crucially, its superior performance on high-round problems highlights the importance of input modeling and the advantage of direct input representations for complex cryptographic structures.

Cryptographic problems, operating within binary variable spaces, can be routinely transformed into Boolean Satisfiability (SAT) problems regarding specific cryptographic conditions like plaintext-ciphertext matching. With the fast development of learning for discrete data, this SAT representation also facilitates the utilization of machine-learning approaches with the hope of automatically capturing patterns and strategies inherent in cryptographic structures in a data-driven manner. Existing neural SAT solvers consistently adopt conjunctive normal form (CNF) for instance representation, which in the cryptographic context can lead to scale explosion and a loss of high-level semantics. In particular, extensively used XOR operations in cryptographic problems can incur an exponential number of clauses. In this paper, we propose a graph structure based on Arithmetic Normal Form (ANF) to efficiently handle the XOR operation bottleneck. Additionally, we design an encoding method for AND operations in these ANF-based graphs, demonstrating improved efficiency over alternative general graph forms for SAT. We then propose CryptoANFNet, a graph learning approach that trains a classifier based on a message-passing scheme to predict plaintext-ciphertext satisfiability. Using ANF-based SAT instances, CryptoANFNet demonstrates superior scalability and can naturally capture higher-order operational information. Empirically, CryptoANFNet achieves a 50x speedup over heuristic solvers and outperforms SOTA learning-based SAT solver NeuroSAT, with 96\% vs. 91\% accuracy on small-scale and 72\% vs. 55\% on large-scale datasets from real encryption algorithms. We also introduce a key-solving algorithm that simplifies ANF-based SAT instances from plaintext and ciphertext, enhancing key decryption accuracy from 76.5\% to 82\% and from 72\% to 75\% for datasets generated from two real encryption algorithms.

Infrared small target detection (IRSTD) is challenging due to low target-background contrast and small target size, leading to missed detections and false alarms (Fas). To address these problems, a novel gradient vector self-learning network (GVSLNet) is proposed. First, the self-learning gradient vector (SLGV) module is designed based on the unique high correlation of infrared small target gradient. Traditional gradient vector field operators cannot update and learn the deep features. SLGV overcomes these limitations by using CNN to adaptively learn and calculate the gradient vector of infrared images, improving the ability to distinguish between targets and backgrounds under complex environments. Then, edge features are encoded into the global-local attention fusion (GLAF) module, which is based on Transformer and dilated convolution. Infrared small target images often display nonlocal self-similarity, where background signals tend to share similar structures. The GLAF module leverages this characteristic to further enhance target intensity while effectively suppressing background noise. The proposed GVSLNet can dynamically calculate gradients based on scene information, improving the detection ability of small targets in complex environments. Experimental results prove that the proposed GVSLNet outperforms state-of-the-art methods on public datasets while maintaining high inference speed.

Tag recommendation in software information sites is a significant task to help developers make distinctions among software objects. Most existing methods usually ignore the semantic information of code snippets in software information sites. To tackle this issue, we regard the code as a semantic enhancement signal, and propose a novel Code-Enhanced fine-grained semantic matching method for Tag Recommendation in software information sites (CETR) to learn the matching score between tags and software objects. In our CETR, code-enhanced semantic interaction is designed to capture fine-grained semantic relevance between tags and software objects. Specifically, code snippets and text descriptions in software objects are first encoded in two specialized ways to capture semantic information from the perspective of natural language and structural language respectively. The semantic interaction is then learned between code snippets and text descriptions. Besides, a hierarchy-aware semantic learning block attentively combines the features from different layers as the final feature to capture semantic information of different levels. Lastly, a pairwise sentence matching computation block is exploited to predict matching probability. Experimental results on four software information site datasets have demonstrated the effectiveness of our proposed CETR for tag recommendation compared with the state-of-the-art methods.

The emergence and rapid progress of the Internet have brought ever-increasing impact on financial domain. How to rapidly and accurately mine the key information from the massive negative financial texts has become one of the key issues for investors and decision Rakers. Aiming at the issue, we propose a sentiment analysis and key entity detection approach based on BERT, which is applied in online financial text mining and public opinion analysis in social media. By using pre-train model, we first study sentiment analysis, and then we consider key entity detection as a sentence matching or Machine Reading Comprehension (MRC) task in different granularity. Among them, we mainly focus on negative sentimental information. We detect the specific entity by using our approach, which is different from traditional Named Entity Recognition (NER). In addition, we also use ensemble learning to improve the performance of proposed approach. Experimental results show that the performance of our approach is generally higher than SVM, LR, NBM, and BERT for two financial sentiment analysis and key entity detection datasets.

Timely and accurate collection and analysis of a financial entity's online text not only helps investors a void risks and optimize investment plans, but also helps the institutions's own public opinion management. However, how to automatically identify the negative text from a large number of online financial text and identify the key entities mainly described by the text is a difficult problem. Most existing studies focuses only on the sentimental polarity of financial texts or only on all the entities that extract them. However, investors and institution may be more focused on key entities in financial texts, rather than just the sentiment of financial texts. Besides, the sentimental polarity of online news is often closely related to key entities which fits well with the idea of multi-task learning. In this paper, we propose a hierarchical multi-task learning framework(HMFF) to not only identify the sentimental polarity of online news but also enhance the feature learning of key entity identification. Experimental results show that our framework not only end-to-end identifies the sentimental polarity of financial text and key entities, but also improves the accuracy of the model.

Tag recommendation in software information sites is a popular way to help developers classify software objects. Existing methods mostly consider tag recommendation as a multi-label classification task, which does not adequately leverage the semantic information of tags themselves. It is observed that the information granularity of tags is from abstract to specific and deep learning models have proven capable of automatically learning the features in different layers of an integrated network with different abstraction degrees. In this paper, we propose TagMatchRec, a deep semantic matching framework for tag recommendation instead of being based on classification. In our framework, multiple layers with different information granularities are directly connected to the output layer aiming at improving the quality of tag recommendation. Moreover, because the abstraction levels of semantic features learned by each layer may be different given different software objects and tags, an attentive deep supervision is introduced so that the dense connections from early layers to the output layer have directly weighted impact on loss function optimization. Comprehensive evaluations are conducted the datasets from four software information sites. The experimental results show that TagMatchRec has achieved better performance compared with the state-of-the-art approaches.