Equipping large language models (LLMs) with external tools has emerged as a promising paradigm for addressing real-world tasks. Nonetheless, with the web-based tool ecosystems rapidly expanding, it is impractical to include all tools within the limited input length of LLMs. To alleviate these issues, researchers have explored incorporating a tool retrieval module to select the most relevant tools or represent tools as unique tokens within LLM parameters. However, most state-of-the-art methods are under transductive settings, assuming all tools have been observed during training. Such a setting deviates from reality as tools on the web are constantly updated and new tools are frequently added to the online ecosystem. When dealing with these unseen tools, which refer to tools not encountered during the training phase, these methods are limited by two key issues, including the large distribution shift and the sensitivity of semantic-only retrieval. To this end, inspired by human cognitive processes of mastering unseen tools through discovering and applying the logical information from prior experience, we introduce a novel Logic-Guided Semantic Bridging framework for inductive tool retrieval, namely, LoSemB, which aims to mine and transfer latent logical information for inductive tool retrieval without costly retraining. Specifically, LoSemB contains a logic-based embedding alignment module to mitigate distribution shifts and a relational augmented retrieval mechanism to overcome the limitations of semantic-only similarity methods. Extensive experiments demonstrate that LoSemB achieves advanced performance in both the inductive and transductive settings.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
LLM-based empathetic dialogue systems enhance agents' emotional support capabilities. Previous approaches primarily relied on Chain-of-Thought (CoT) prompting to extract key dialogue cues and further strengthened the agent's sensitivity to these signals through supervised fine-tuning. However, such methods overly depend on the information extraction capability of LLMs, leading to unstable reasoning and limited interpretability. To simultaneously improve an agent's ability to proactively explore solutions through rational reasoning while attending to users' sensitive emotions via empathetic understanding, we propose Neuro-Sym Supporter, a hybrid decision-making emotional support agent that integrates symbolic reasoning with deep learning. This model combines rational inference with emotional empathy, enabling the agent to generate supportive responses that balance logic and emotion. Specifically, we introduce Sym-Mind, a differentiable logic-based reasoning framework for emotional support strategy selection, which unifies interpretability with stable performance. Experimental results on public datasets demonstrate that our approach consistently outperforms multiple competitive baselines in both automatic and human evaluations, validating its effectiveness.
Large language models (LLMs) show promising performance on small-scale graph reasoning tasks but fail when handling real-world graphs with complex queries. This phenomenon arises from LLMs' working memory constraints, which result in their inability to retain long-range graph topology over extended contexts while sustaining coherent multi-step reasoning. However, real-world graphs are often structurally complex, such as Web, Transportation, Social, and Citation networks. To address these limitations, we propose GraphCogent, a collaborative agent framework inspired by human Working Memory Model that decomposes graph reasoning into specialized cognitive processes: sense, buffer, and execute. The framework consists of three modules: Sensory Module standardizes diverse graph text representations via subgraph sampling, Buffer Module integrates and indexes graph data across multiple formats, and Execution Module combines tool calling and tool creation for efficient reasoning. We also introduce Graph4real, a comprehensive benchmark that contains four domains of real-world graphs (Web, Transportation, Social, and Citation) to evaluate LLMs' graph reasoning capabilities. Our Graph4real covers 21 different graph reasoning tasks, categorized into three types (Structural Querying, Algorithmic Reasoning, and Predictive Modeling tasks), with graph scales up to 10 times larger than existing benchmarks. Experiments show that Llama3.1-8B based GraphCogent achieves a 50% improvement over massive-scale LLMs like DeepSeek-R1 (671B). Compared to state-of-the-art code-based baseline, our framework outperforms by 20% in accuracy while reducing token usage by 80% for in-toolset tasks and 30% for out-toolset tasks.
Dual History Enhancement with Hybrid Hypergraph-Graph Networks for Temporal Knowledge Graph Reasoning
PDF ↗Temporal Knowledge Graph (TKG) reasoning seeks to predict future events by analyzing historical data, where the effective leverage of both local and global historical facts proves crucial. Existing approaches employ graph neural networks (GNNs) and recurrent neural networks (RNNs) for local evolution patterns, complemented by statistical methods to enhance attention to global facts, demonstrating efficient predictive capabilities. However, traditional GNNs, constrained by their low-order neighborhood aggregation design, inherently fail to model potential high-order dependencies among facts. Furthermore, existing global history modeling approaches may introduce irrelevant historical information that interferes with prediction tasks. To address these limitations, we propose a Dual History-aware HyperGraph Network for TKG reasoning, namely DHHGN. Specifically, for local history modeling, we design a hybrid hypergraph-graph joint recurrent convolution module that simultaneously captures low-order neighborhood information and high-order interaction patterns among entities, employing a gating mechanism to adaptively blend their contributions. For global history modeling, we propose a dual history enhancement module that amplifies attention on pivotal historical facts while ensuring holistic integration of all historical contexts. Extensive experiments on four public benchmarks validate that DualHist-HGN consistently outperforms existing state-of-the-art methods across TKG reasoning tasks.
With the rapid growth of multi-modal content on the Web, robust vision-language models are essential for semantic understanding and classification of web images under diverse and dynamic contexts, supporting Web applications such as multimedia search and recommendation. Prompt learning has proven effective for enhancing vision-language models in semantic image classification tasks. However, previous methods often suffer from poor generalization: the learned prompts tend to overfit the base classes seen during training, leading to poor performance on unseen classes and under distribution shifts. This issue is especially challenging in Web-scale data, where new classes emerge and distributions shift dynamically. To address these limitations, we propose PLIKD, a novel prompt learning method that integrates instance-aware knowledge distillation for robust Web-scale semantic image classification. Specifically, PLIKD introduces an instance-aware knowledge extraction module, which leverages multi-modal large language models through a step-by-step strategy to extract external knowledge for each image instance. To incorporate this extracted knowledge, PLIKD further introduces an instance-aware knowledge distillation module, which consists of two key steps: (1) a dual-teacher strategy for robust and informative knowledge distillation, and (2) fine-grained cross-modal alignment via Smooth and Sparse Optimal Transport. Extensive experiments demonstrate that PLIKD significantly improves generalization to both seen and unseen classes, and remains robust under distribution shifts, outperforming existing state-of-the-art methods on Web-scale semantic image classification.
The multimodal Chinese idiom reading comprehension task aims to select the most appropriate idiom from a candidate list via the given text and image. This poses a significant challenge for the model to comprehend each Chinese idiom accurately. Existing multimodal Chinese idiom reading comprehension methods primarily focus on aligning contextual text and images, while overlooking two key attributes of Chinese idioms.(1) There is a discrepancy between the literal and metaphorical meanings of Chinese idioms. (2) The same Chinese idiom has different meanings in different scenarios, which requires targeted understanding by experts who specialize in different fields. To address the above challenges, we rethink the solution to the multimodal idiom reading comprehension task from a metaphorical perspective and propose a framework named MePe. Firstly, we propose a literal metaphorical semantic graph that systematically transforms the implicit discrepancy between the literal and metaphorical meanings of Chinese idioms into structured explicit relationships, thereby making metaphorical meanings more understandable. Then, we propose a mixture of idiom experts consisting of a literal idiom expert and a metaphorical idiom expert. Through division of labor and collaboration among these experts, we achieve an understanding of the dual meanings of Chinese idioms across different scenarios. Finally, we employ the maximum mean discrepancy to adjust the variance between the literal and metaphorical semantic features of Chinese idioms. By mapping these features onto a shared reproducing kernel Hilbert space, the model can better distinguish between the two based on contextual clues. Extensive experiments demonstrate that MePe achieves state-of-the-art performance on the MChIRC dataset.
Pseudo-Alignment is a pervasive challenge in many large language models for time series (LLM4TS) models, often causing them to underperform compared to linear models or randomly initialised backbones. However, there is limited discussion in the community for the reasons that pseudo-alignment occurs. In this work, we conduct a thorough investigation into the root causes of pseudo-alignment in LLM4TS and build a connection of pseudo-alignment to the cone effect in LLM. We demonstrate that pseudo-alignment arises from the interplay of cone effect within pretrained LLM components and the intrinsically low-dimensional manifold of time-series data. In addition, we also introduce TimeSUP, a novel technique designed to mitigate this issue and improve forecast performance in existing LLM4TS approaches. TimeSUP addresses this by increasing the time series manifold to more closely match the intrinsic dimension of language embeddings, allowing the model to distinguish temporal signals clearly while still capturing shared structures across modalities. As a result, representations for time and language tokens remain distinct yet exhibit high cosine similarity, signifying that the model preserves each modality's unique features while learning their commonalities in a unified embedding space. Empirically, TimeSUP consistently outperforms state-of-the-art LLM4TS methods and other lightweight baselines on long-term forecasting performance. Furthermore, it can be seamlessly integrated into four existing LLM4TS pipelines and delivers significant improvements in forecasting performance.
Web-scale platforms and online services rely on log-based anomaly detection to safeguard availability, latency SLOs, and user experience. In real-world web interactions, system logs often exhibit irregular temporal intervals, bursty densities, and heterogeneous semantics, which pose significant challenges for log anomaly detection. Existing methods such as LSTM and Transformer assume a fixed input window, which conflicts with the inherently irregular nature of system logs. Moreover, most prior works build a single-view representation, overlooking the multi-relational nature of logs. To overcome these challenges, we propose DyLogNet, a dynamic multi-relational graph framework for log anomaly detection. Specifically, this framework constructs a density-aware dynamic graph with variable-length windows, and represents logs from three relational perspectives: temporal co-occurrence, semantic similarity, and anomaly tendency. Next, we design a cross-layer attention mechanism that integrates heterogeneous structures to highlight the most relevant relations and enhance event representations. Furthermore, a cross-snapshot memory injection module updates global memory through a recurrent unit and injects it into current graph representations via an affine transformation, enabling temporal continuity. Experiments on three public log datasets demonstrate that DyLogNet outperforms state-of-the-art methods, especially in few-shot scenarios.
Multi-domain knowledge graph completion (MKGC) aims to improve missing triple prediction in a target KG by transferring knowledge from other support KGs. Existing methods typically enforce consistency constraints on equivalent entities across KGs to transfer knowledge, which risks suppressing domain-specific contextual information of entities. This design can also compromise entity representation information from all KG domains, impeding performance improvements, especially in low-resource data scenarios. To address this, we pioneer a generation-based paradigm for MKGC and propose DMKGC, a conditional diffusion-guided knowledge transfer framework. Our key insight is to treat each KG as a partial view of the entity entire information, and generate informative domain-general entity embeddings through diffusion models conditioned on support KGs. Particularly, we first initialize domain-agnostic entity embeddings as prior entity embeddings, and then encode them within individual KGs. Afterward, we fuse equivalent entities from support KGs as the conditional diffusion generation guidance. We leverage the prior entity embeddings as the proxy generation objective, which ensures this conditional generation to be unbiased towards any conditioned KGs. Simultaneously, we also train the generated embeddings to be predictive across KGs, thus preserving domain-specific information. Extensive experiments on 14 KGs in 3 benchmarks demonstrate a 4.3% average MRR improvement in tail entity prediction over state-of-the-art methods, with sustained gains in low-resource data settings.
Revealing the underlying causal mechanisms in the real world is crucial for scientific and technological progress. Despite notable advances in recent decades, the lack of high-quality data and the reliance of traditional causal discovery algorithms (TCDA) on the assumption of no latent confounders, as well as their tendency to overlook the precise semantics of latent variables, have long been major obstacles to the broader application of causal discovery. To address this issue, we propose a novel causal modeling framework, TLVD, which integrates the metadata-based reasoning capabilities of large language models (LLMs) with the data-driven modeling capabilities of TCDA for inferring latent variables and their semantics. Specifically, we first employ a data-driven approach to construct a causal graph that incorporates latent variables. Then, we employ multi-LLM collaboration for latent variable inference, modeling this process as a game with incomplete information and seeking its Bayesian Nash Equilibrium (BNE) to infer the possible specific latent variables. Finally, to validate the inferred latent variables across multiple real-world web-based data sources, we leverage LLMs for evidence exploration to ensure traceability. We comprehensively evaluate TLVD on three de-identified real patient datasets provided by a hospital and two benchmark datasets. Extensive experimental results confirm the effectiveness and reliability of TLVD, with average improvements of 32.67% in Acc, 62.21% in CAcc, and 26.72% in ECit across the five datasets.
CausalSKyHop: Knowledge-Aware Causal Explanation of Dynamic GNNs via Higher-Order Semantic Reasoning
Dynamic Graph Neural Networks (DyGNNs) are widely used to model web-scale semantic-rich graph data (e.g., social networks, knowledge graphs), but their inability to explain predictions ground-ed in structured knowledge remains a challenge, especially when predictions rely on complex higher-order substructures. We propose CausalSKyHop, a semantic- and knowledge-aware framework that explains DyGNNs by uncovering causal higher-order patterns in evolving knowledge structures. To model the semantic fabric of the graph, CausalSKyHop incorporates a Higher-Order Structural Causal Model to capture multi-node knowledge dependencies, and uses contrastive learning to isolate semantically-meaningful causal relationships from spurious ones. A dynamic correlation module further separates persistent knowledge from evolving semantic contexts. Through knowledge-infused, structure-aware variational graph autoencoders, our method produces interpretable causal subgraphs that capture the dynamic flow of knowledge and semantics. Experimental evaluations on multiple web and knowledge-rich graph benchmarks demonstrate that CausalSKyHop consistently outperforms state-of-the-art explainable DyGNNs, achieving notable improvements in both explanation fidelity and downstream prediction accuracy. A detailed case study further illustrates how our method uncovers stable, semantically coherent causal pathways—in contrast to the fragmented explanations of baseline methods—providing intuitive evidence for its superior interpretability. This work establishes the critical role of explicit semantic and knowledge integration through higher-order causal reasoning for building transparent and trustworthy DyGNNs on the web.
The explosive growth of multimodal web data demands communication that transmits meaning rather than raw bits. Existing semantic-communication systems often fail under noise, missing modalities, and distribution shifts because they optimize surface features instead of modality-invariant knowledge. We present Grasp, a knowledge-centric framework for cross-modal communication. Grasp segments streams into semantic blocks and builds a graph over them; a lightweight Graph Neural Networks (GNN) produces schedulable, importance-weighted representations. At its core is knowledge purification : we minimize a conditional mutual information upper bound to perform a three-way disentanglement—strongly related, weakly related, and task-irrelevant components—so that only essential semantics are transmitted while non-essential factors are suppressed. To maintain synchrony, we introduce one-to-two temporal contrastive learning to achieve triple alignment of video, audio, and text despite sampling asynchrony. For efficient transmission, Grasp uses a cross-modal shared vector-quantization codebook—a discrete knowledge codebook —updated by multimodal attention. At the receiver, a soft-recovery mechanism leverages this shared knowledge to robustly reconstruct semantics under low signal-to-noise ratio (SNR) or missing modalities, yielding graceful degradation. Across web tasks—including cross-modal retrieval and missing-modality inference—Grasp improves knowledge consistency, semantic fidelity, and downstream performance over strong baselines while maintaining low latency. These results show that communication structured around purified knowledge is key to building robust, semantic-aware systems for the modern web.
Text classification is a crucial and fundamental task in web content mining. Compared with the previous learning paradigm of pre-training and fine-tuning by cross entropy loss, the recently proposed supervised contrastive learning approach has received tremendous attention due to its powerful feature learning capability and robustness. Although several studies have incorporated this technique for text classification, some limitations remain. First, many text datasets are imbalanced, and the learning mechanism of supervised contrastive learning is sensitive to data imbalance, which may harm the model's performance. Moreover, these models leverage separate classification branches with cross entropy and supervised contrastive learning branches without explicit mutual guidance. To this end, we propose a novel model named SharpReCL for imbalanced text classification tasks. First, we obtain the prototype vector of each class in the balanced classification branch to act as a representation of each class. Then, by further explicitly leveraging the prototype vectors, we construct a proper and sufficient target sample set with the same size for each class to perform the supervised contrastive learning procedure. The empirical results show the effectiveness of our model, which even outperforms popular large language models across several datasets. Our code is available https://github.com/KEAML-JLU/SharpReCL
As knowledge and semantics on the web grow increasingly complex, enhancing Large Language Models (LLMs)' comprehension and reasoning capabilities has become particularly important. Chain-of-Thought (CoT) prompting has been shown to enhance the reasoning capabilities of LLMs. However, it still falls short on logical reasoning tasks that rely on symbolic expressions and strict deductive rules. Neuro-symbolic methods address this gap by enforcing formal correctness through external solvers. Yet these solvers are highly format-sensitive, and small instabilities in model outputs can lead to frequent processing failures. The LLM-driven approaches avoid parsing brittleness, but they lack structured representations and process-level error-correction mechanisms. To further enhance the logical reasoning capabilities of LLMs, we propose MatrixCoT, a structured CoT framework with a matrix-based plan. Specifically, we normalize and type natural language expressions and attach explicit citation fields, and introduce a matrix-based planning method to preserve global relations among steps. The plan thus becomes a verifiable artifact and execution becomes more stable. For verification, we also add a feedback-driven replanning mechanism. Under semantic-equivalence constraints, it identifies omissions and defects, rewrites and compresses the dependency matrix, and produces a more trustworthy final answer. Experiments on five logical-reasoning benchmarks and five LLMs show that, without relying on external solvers, MatrixCoT enhances both the robustness and interpretability of LLMs when tackling complex symbolic reasoning tasks, while maintaining competitive performance.
Multi-modal knowledge graphs (MMKGs) enrich traditional knowledge graphs by incorporating heterogeneous modalities such as textual descriptions and visual content, offering complementary semantic cues for knowledge reasoning. However, existing approaches often overlook the structural dependencies within each modality, apply static or coarse-grained fusion strategies, and insufficiently model relational semantics. We propose a Multi-Granularity Multi-Modal Knowledge Graph Representation Learning Method via Subgraph-aware Adaptive Fusion and Hierarchical Relation Modeling (SAFER ), which implement multi-modal knowledge representation through adaptive fusion of multi-granularity information such as multi-modal semantics, knowledge structures and relations. SAFER explicitly constructs modality-specific subgraphs and employs structure-aware graph attention networks to effectively capture intra-modal structural dependencies. We propose an adaptive multi-modal fusion mechanism, which aggregates modality-specific embeddings at the semantic level by dynamically assigning entity-specific modality weights. We design a two-stage multi-granularity knowledge relation modeling strategy, which utilizes a structure-aware multi-modal adaptive pre-fusion to preserve topological information and a relation-aware graph attention network (RGAT) post-fusion to encode relational semantics. Extensive experiments on several benchmark datasets demonstrate that the proposed SAFER significantly outperforms competitive baselines on link prediction and relation reasoning tasks.
Knowledge Graph Question Answering (KGQA) aims to answer natural language questions by reasoning over structured knowledge graphs (KGs). While large language models (LLMs) have advanced KGQA through their strong reasoning capabilities, existing methods continue to struggle to fully exploit both the rich knowledge encoded in KGs and the reasoning capabilities of LLMs, particularly in complex scenarios. They often assume complete KG coverage and lack mechanisms to judge when external information is needed, and their reasoning remains locally myopic, failing to maintain coherent multi-step planning, leading to reasoning failures even when relevant knowledge exists. We propose Graph-RFT, a novel two-stage reinforcement fine-tuning KGQA framework with a ''plan–KGsearch–and–Websearch–during–think'' paradigm, that enables LLMs to perform autonomous planning and adaptive retrieval scheduling across KG and web sources under incomplete knowledge conditions. Graph-RFT introduces a chain-of-thought (CoT) fine-tuning method with a customized plan–retrieval dataset activates structured reasoning and resolves the GRPO cold-start problem. It then introduces a novel plan–retrieval guided reinforcement learning process integrates explicit planning and retrieval actions with a multi-reward design, enabling coverage-aware retrieval scheduling. It employs a Cartesian-inspired planning module to decompose complex questions into ordered sub-questions, and logical expression to guide tool invocation for globally consistent multi-step reasoning. This reasoning–retrieval process is optimized with a multi-reward combining outcome and retrieval-specific signals, enabling the model to learn when and how to combine KG and web retrieval effectively. Experiments on multiple KGQA benchmarks demonstrate that Graph-RFT achieves superior performance over strong baselines, even with smaller LLM backbones, and substantially improves complex question decomposition, factual coverage, and tool coordination.
Semantic understanding of popularity bias is a crucial yet underexplored challenge in recommender systems, where popular items are often favored at the expense of niche content. Most existing debiasing methods treat the semantic understanding of popularity bias as a matter of diversity enhancement or long-tail coverage, neglecting the deeper semantic layer that embodies the causal origins of the bias itself. Consequently, such shallow interpretations limit both their debiasing effectiveness and recommendation accuracy. In this paper, we propose FairLRM, a novel framework that bridges the gap in the semantic understanding of popularity bias with Recommendation via Large Language Model (RecLLM). FairLRM decomposes popularity bias into item-side and user-side components, using structured instruction-based prompts to enhance the model's comprehension of both global item distributions and individual user preferences. Unlike traditional methods that rely on surface-level features such as ''diversity'' or ''debiasing'', FairLRM improves the model's ability to semantically interpret and address the underlying bias. Through empirical evaluation, we demonstrate that FairLRM enhances fairness and recommendation accuracy through a trustworthy, semantically grounded treatment of popularity bias. The source code is shown in https://github.com/LuoRenqiang/FairLRM.
Temporal Knowledge Graph Completion (TKGC) aims to predict missing entities or relations based on historical facts, thereby facilitating the understanding of dynamic system evolution and supporting downstream reasoning tasks. However, existing methods predominantly focus on modeling sequential and structural dependencies, often overlooking the rich semantic information embedded in entities and relations, as well as the higher-order interactions among them, which limits their ability to handle complex, evolving scenarios effectively. To address these limitations, we propose DSTAG, a novel TKGC approach based on a semantic tag-enhanced dual-graph convolutional network. Our method leverages large language models to generate contextualized semantic multi-tags for both entities and relations (e.g., ''political event,'' ''economic activity''), thereby enriching their semantic representations. Furthermore, we introduce a semantic tag representation mechanism that captures higher-order dependencies during the aggregation and propagation of semantic tag information across graphs. DSTAG adopts a dual-graph convolutional network architecture, where the relation graph convolution extracts semantic features between temporal relationships and injects this information into the entity graph convolution, enabling joint modeling of entities and relations. We evaluate DSTAG on three widely used TKG benchmarks: ICEWS14, ICEWS18, and ICEWS05-15. Experimental results show that DSTAG achieves substantial MRR improvements over state-of-the-art baselines by 8.64%, 9.81% and 4.56%, respectively.
Logical reasoning with large language models (LLMs) has received growing attention. One mainstream approach translates natural language into formal logic and then applies symbolic solvers for deduction. While effective in many tasks, these LLM-based translators often fail to generate consistent symbolic representations when the same concept appears in different linguistic forms. Such inconsistencies break logical coherence and lead to solver errors. However, most existing benchmarks lack this type of linguistic variation, which frequently occurs in real-world text, leaving the problem underexplored. To address this gap, we present SoLT, a benchmark that systematically rewrites reasoning datasets into diverse yet logically equivalent forms across multiple levels. Beyond evaluation, SoLT also provides a general method to enrich any dataset with linguistic diversity while preserving both meaning and logic. To further enhance the stability of LLM-based reasoning, we propose MenTaL, which explicitly guides models to build a concept–symbol mapping table during translation. By linking equivalent expressions to shared symbols, MenTaL maintains consistency and mitigates symbol drift. Experiments on SoLT demonstrate that LLMs indeed suffer from inconsistent symbol mapping under linguistic variation, leading to significant drops in reasoning accuracy. Meanwhile, applying MenTaL brings clear and stable performance improvements across diverse inputs. Overall, our findings reveal that overlooking linguistic diversity hides key weaknesses in LLM-based translators, and our work offers a step toward more reliable logical reasoning in varied real-world scenarios. Our code is available at https://github.com/wufeiwuwoshihua/LinguDiver.
While Large Language Models (LLMs) are increasingly deployed in Web applications such as search, dialogue, and recommendation systems, their reliance on large-scale Web data raises serious privacy concerns, particularly the risk of memorizing sensitive content. Existing Membership Inference Attacks (MIA) rely heavily on the surface form of inputs, rendering them ineffective against semantically preserved but structurally altered samples. This methodological weakness results in widespread false negatives and compromises the integrity of privacy evaluations in large-scale Web corpora. To address this limitation, we propose Adversarial Semantic Membership Inference Attack (ASMIA). ASMIA enhances MIA effectiveness by generating semantically diverse adversarial samples, extracting multi-layer attention features from the target model, and training a contrastive classifier that leverages similarity metrics and logarithmic probabilities to distinguish members from non-members. Experiments on LLMs trained with Wikipedia, a representative large-scale Web corpus, demonstrate that ASMIA significantly outperforms existing methods, highlighting the value of semantic perturbations and attention patterns in detecting training data leakage.