Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of ''Spontaneous Multilingual Alignment''. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.
论文检索
输入标题、作者或关键词，从 1,144 篇学术成果中精准定位
Improving Implicit Discourse Relation Recognition with Natural Language Explanations from LLMs
PDF ↗Implicit Discourse Relation Recognition (IDRR) remains a challenging task due to the requirement for deep semantic understanding in the absence of explicit discourse markers. A further limitation is that existing methods only predict relations without providing any supporting explanations. Recent advances in large language models (LLMs) have shown strong reasoning capabilities in both deep language understanding and natural language explanation generation. In this work, we propose a simple yet effective approach to distill the reasoning capabilities of LLMs into lightweight IDRR models to improve both performance and interpretability. Specifically, we first prompt an LLM to generate explanations for each training instance conditioned on its gold label. Then, we introduce a novel classification-generation framework that jointly performs relation prediction and explanation generation, and train it with the additional supervision of LLM-generated explanations. Our framework is plug-and-play, enabling easy integration with most existing IDRR models. Experimental results on PDTB demonstrate that our approach significantly improves IDRR performance, while human evaluation further confirms that the generated explanations enhance model interpretability. Furthermore, we validate the generality of our approach on sentiment classification and natural language inference.
Enhancing Meme Emotion Understanding with Multi-Level Modality Enhancement and Dual-Stage Modal Fusion
PDF ↗With the rapid rise of social media and Internet culture, memes have become a popular medium for expressing emotional tendencies. This has sparked growing interest in Meme Emotion Understanding (MEU), which aims to classify the emotional intent behind memes by leveraging their multimodal contents. While existing efforts have achieved promising results, two major challenges remain: (1) a lack of fine-grained multimodal fusion strategies, and (2) insufficient mining of memes' implicit meanings and background knowledge. To address these challenges, we propose MemoDetector, a novel framework for advancing MEU. First, we introduce a four-step textual enhancement module that utilizes the rich knowledge and reasoning capabilities of Multimodal Large Language Models (MLLMs) to progressively infer and extract implicit and contextual insights from memes. These enhanced texts significantly enrich the original meme contents and provide valuable guidance for downstream classification. Next, we design a dual-stage modal fusion strategy: the first stage performs shallow fusion on raw meme image and text, while the second stage deeply integrates the enhanced visual and textual features. This hierarchical fusion enables the model to better capture nuanced cross-modal emotional cues. Experiments on two datasets, MET-MEME and MOOD, demonstrate that our method consistently outperforms state-of-the-art baselines. Specifically, MemoDetector improves F1 scores by 4.3% on MET-MEME and 3.4% on MOOD. Further ablation studies and in-depth analyses validate the effectiveness and robustness of our approach, highlighting its strong potential for advancing MEU.
Deploying language models often requires navigating accuracy vs. performance trade-offs to meet latency constraints while preserving utility. Traditional model distillation reduces size but incurs substantial costs through training separate models. We introduce ModularStarEncoder (MoSE), a 1-billion-parameter multi-exit encoder for code retrieval and classification that employs a novel Self-Distillation mechanism. This approach significantly enhances lower-layer representations, enabling flexible deployment of different model portions with favorable performance trade-offs. Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads, where higher layers guide earlier ones during training, thereby improving intermediate representations at minimal additional cost. We further enhance MoSE with a repository-level contextual loss that maximizes training context window utilization. Additionally, we release a new dataset created through code translation that extends text-to code benchmarks with cross-language code-to-code pairs. Evaluations demonstrate the effectiveness of Self-Distillation as a principled approach to trading inference cost for accuracy across various code understanding tasks.
Large language models have demonstrated impressive performance on natural language tasks, but their decision-making processes remain opaque. Existing explanation methods either suffer from limited faithfulness to the model's reasoning or produce explanations that are difficult for humans to understand. To address these challenges, we propose ProtoSurE, a novel prototype-based surrogate framework that provides faithful and understandable explanations for LLMs. ProtoSurE trains an interpretable-by-design surrogate model that aligns with the target LLM while utilizing sentence-level prototypes as understandable concepts. Extensive experiments show that ProtoSurE consistently outperforms state-of-the-art explanation methods across diverse LLMs and datasets. Importantly, ProtoSurE demonstrates strong data efficiency, requiring relatively few training examples to achieve good performance, making it practical for real-world applications.
MAU-GPT: Enhancing Multi-type Industrial Anomaly Understanding via Anomaly-aware and Generalist Experts Adaptation
PDF ↗As industrial manufacturing scales, automating fine-grained product image analysis has become critical for quality control. However, existing approaches are hindered by limited dataset coverage and poor model generalization across diverse and complex anomaly patterns. To address these challenges, we introduce MAU-Set, a comprehensive dataset for Multi-type industrial Anomaly Understanding. It spans multiple industrial domains and features a hierarchical task structure, ranging from binary classification to complex reasoning. Alongside this dataset, we establish a rigorous evaluation protocol to facilitate fair and comprehensive model assessment. Building upon this foundation, we further present MAU-GPT, a domain-adapted multimodal large model specifically designed for industrial anomaly understanding. It incorporates a novel AMoE-LoRA mechanism that unifies anomaly-aware and generalist experts adaptation, enhancing both understanding and reasoning across diverse defect classes. Extensive experiments show that MAU-GPT consistently outperforms prior state-of-the-art methods across all domains, demonstrating strong potential for scalable and automated industrial inspection.
DWTSG: Parameter-Efficient Fine-Tuning of Large Pre-trained Models via Discrete Wavelet Transform and Subband Guidance
PDF ↗Fully fine-tuning large pre-trained models for each downstream task is impractical due to prohibitive memory, computation, and storage costs. Although parameter-efficient fine-tuning (PEFT) methods address this issue, leading methods like LoRA still exhibit linear scaling of trainable parameters with hidden size. Recent studies have explored PEFT in the frequency domain to reduce computational costs by employing fast Fourier transform and discrete cosine transform with sparse frequency selection. These methods rely on global frequency representations that lack spatial locality and disperse energy across the domain. As a result, sparse coefficient selection struggles to preserve fine-grained structural information and often introduces artifacts such as ringing near boundaries. To address these limitations, we propose DWTSG, a novel PEFT framework based on discrete wavelet transform (DWT) and subband guidance. DWTSG decomposes pre-trained weights into four wavelet subbands that jointly encode global context and local details. It fine-tunes only the most informative coefficients in each subband through an energy-based selection strategy that prioritizes coefficients based on their individual importance and interactions. Finally, inverse DWT reconstructs the updated weights, enabling efficient and precise adaptation. Extensive experiments on natural language understanding, commonsense reasoning, and image classification demonstrate that DWTSG outperforms existing PEFT methods, achieving superior performance and higher parameter efficiency.
Catastrophic forgetting is a longstanding challenge in continual learning, where models lose knowledge from earlier tasks when learning new ones. While various mitigation strategies have been proposed for Multi-Layer Perceptrons (MLPs), recent architectural advances like Kolmogorov-Arnold Networks (KANs) have been suggested to offer intrinsic resistance to forgetting by leveraging localized spline-based activations. However, the practical behavior of KANs under continual learning remains unclear, and their limitations are not well understood. To address this, we present a comprehensive study of catastrophic forgetting in KANs and develop a theoretical framework that links forgetting to activation support overlap and intrinsic data dimension. We validate these analyses through systematic experiments on synthetic and vision tasks, measuring forgetting dynamics under varying model configurations and data complexity. Further, we introduce KAN-LoRA, a novel adapter design for parameter-efficient continual fine-tuning of language models, and evaluate its effectiveness in knowledge editing tasks. Our findings reveal that while KANs exhibit promising retention in low-dimensional algorithmic settings, they remain vulnerable to forgetting in high-dimensional domains such as image classification and language modeling. These results advance the understanding of KANs’ strengths and limitations, offering practical insights for continual learning system design.
We investigate a binary classification problem in an edge intelligence system where false negatives are more costly than false positives. The system features a compact, locally deployed model, supplemented by a larger, remote model that is accessible via the network, albeit at an offloading cost. For each sample, our system first uses the locally deployed model for inference. Based on the output of the local model, the sample may be offloaded to the remote model. This work aims to understand the fundamental trade-off between classification accuracy and the offloading costs within such a hierarchical inference (HI) system. To optimise this system, we propose an online learning framework that continuously adapts a pair of thresholds on the local model's confidence scores. These thresholds determine the prediction of the local model and whether a sample is classified locally or offloaded to the remote model. We present a closed-form solution for the setting where the local model is calibrated. For the more general case of uncalibrated models, we introduce H2T2, an online two-threshold hierarchical inference policy, and prove it achieves sublinear regret. H2T2 is model-agnostic, requires no training, and learns during the inference phase using limited feedback. Simulations on real-world datasets show that H2T2 consistently outperforms naive and single-threshold HI policies, sometimes even surpassing single-threshold offline optima. The policy also demonstrates robustness to distribution shifts and adapts effectively to mismatched classifiers.
LLM-Enhanced Energy Contrastive Learning for Out-of-Distribution Detection in Text-Attributed Graphs
PDF ↗Text-attributed graphs, where nodes are enriched with textual attributes, have become a powerful tool for modeling real-world networks such as citation, social, and transaction networks. However, existing methods for learning from these graphs often assume that the distributions of training and testing data are consistent. This assumption leads to significant performance degradation when faced with out-of-distribution (OOD) data. In this paper, we address the challenge of node-level OOD detection in text-attributed graphs, with the goal of maintaining accurate node classification while simultaneously identifying OOD nodes. We propose a novel approach, LLM-Enhanced Energy Contrastive Learning for Out-of-Distribution Detection in Text-Attributed Graphs (LECT), which integrates large language models (LLMs) and energy-based contrastive learning. The proposed method involves generating high-quality OOD samples by leveraging the semantic understanding and contextual knowledge of LLMs to create dependency-aware pseudo-OOD nodes, and applying contrastive learning based on energy functions to distinguish between in-distribution (IND) and OOD nodes. The effectiveness of our method is demonstrated through extensive experiments on six benchmark datasets, where our method consistently outperforms state-of-the-art baselines, achieving both high classification accuracy and robust OOD detection capabilities.
Efficiently fine-tuning pre-trained models for downstream tasks is a key challenge in the era of foundation models. Parameter-efficient fine-tuning (PEFT) presents a promising solution, achieving performance comparable to full fine-tuning by updating only a small number of adaptation weights per layer. Traditional PEFT methods typically rely on a single expert, where the adaptation weight is a low-rank matrix. However, for complex tasks, the data's inherent diversity poses a significant challenge for such models, as a single adaptation weight cannot adequately capture the features of all samples. To address this limitation, we explore how to integrate multiple small adaptation experts into a compact structure to defeat a large adapter. Specifically, we propose Tucker Adaptation (TuckA), a method with four key properties: (i) We use Tucker decomposition to create a compact 3D tensor where each slice naturally serves as an expert. The low-rank nature of this decomposition ensures that the number of parameters scales efficiently as more experts are added. (ii) We introduce a hierarchical strategy that organizes these experts into groups at different granularities, allowing the model to capture both local and global data patterns. (iii) We develop an efficient batch-level routing mechanism, which reduces the router's parameter size by a factor of L compared to routing at every adapted layer (where L is the number of adapted layers) (iv) We propose data-aware initialization to achieve loss-free expert load balancing based on theoretical analysis. Extensive experiments on benchmarks in natural language understanding, image classification, and mathematical reasoning speak to the efficacy of TuckA, offering a new and effective solution to the PEFT problem.
Existing feature engineering methods based on large language models (LLMs) have not yet been applied to multi-label learning tasks. They lack the ability to model complex label dependencies and are not specifically adapted to the characteristics of multi-label tasks. To address the above issues, we propose Feature Engineering Automation for Multi-Label Learning (FEAML), an automated feature engineering method for multi-label classification which leverages the code generation capabilities of LLMs. By utilizing metadata and label co-occurrence matrices, LLMs are guided to understand the relationships between data features and task objectives, based on which high-quality features are generated.The newly generated features are evaluated in terms of model accuracy to assess their effectiveness, while Pearson correlation coefficients are used to detect redundancy. FEAML further incorporates the evaluation results as feedback to drive LLMs to continuously optimize code generation in subsequent iterations. By integrating LLMs with a feedback mechanism, FEAML realizes an efficient, interpretable and self-improving feature engineering paradigm. Empirical results on various multi-label datasets demonstrate that our FEAML outperforms other feature engineering methods.
Medical vision–language pretraining typically relies on static image–text pairs, overlooking temporal cues vital for understanding clinical progression. This limits model sensitivity to evolving semantics and reduces their effectiveness in real-world clinical reasoning. To address this challenge, we propose TAMM—a temporal alignment framework that leverages weak but semantically rich supervision from large language models (LLMs). Given temporally adjacent clinical reports, LLMs automatically generate (i) coarse-grained trend labels (e.g., improving or worsening), and (ii) fine-grained rationales explaining the supporting clinical evidence. These complementary signals inject temporal semantics without requiring manual annotation, and guide vision–language representation learning to capture trend-sensitive cross-modal alignment and rationale-grounded coherence. Experiments on multiple medical benchmarks demonstrate that TAMM improves retrieval and classification performance while yielding more interpretable, temporally consistent embeddings. Our results highlight the potential of leveraging LLM-derived supervision to equip vision–language models with temporal awareness critical for clinical applications.
We present NoReGeo, a novel benchmark designed to evaluate the intrinsic geometric understanding of large language models (LLMs) without relying on reasoning or algebraic computation. Unlike existing benchmarks that primarily assess models' proficiency in reasoning-based geometry-where solutions are derived using algebraic methods-NoReGeo focuses on evaluating whether LLMs can inherently encode spatial relationships and recognize geometric properties directly. Our benchmark comprises 2,500 trivial geometric problems spanning 25 categories, each carefully crafted to be solvable purely through native geometric understanding, assuming known object locations. We assess a range of state-of-the-art models on NoReGeo, including frontier models like GPT-4, observing that even the most advanced systems achieve an overall maximum of 65% accuracy in binary classification tasks. Further, our ablation experiments demonstrate that such geometric understanding does not emerge through fine-tuning alone, indicating that effective training for geometric comprehension requires a specialized approach from the outset. Our findings highlight a significant gap in current LLMs' ability to natively grasp geometric concepts, providing a foundation for future research toward models with true geometric cognition.
Large Language Models (LLMs) hold rich implicit knowledge and powerful transferability. In this paper, we explore the combination of LLMs with the human skeleton to perform action classification and description. However, when treating LLM as a recognizer, two questions arise: 1) How can LLMs understand the skeleton? 2) How can LLMs distinguish among actions? To address these problems, we introduce a novel paradigm named learning Skeleton representation with visual-motion knowledge for Action Recognition (SUGAR). In our pipeline, we first utilize off-the-shelf large-scale video models as a knowledge base to generate visual, motion information related to actions. Then, we propose to supervise skeleton learning through this prior knowledge to yield discrete representations. Finally, we use the LLM with untouched pre-training weights to understand these representations and generate the desired action targets and descriptions. Notably, we present a Temporal Query Projection (TQP) module to continuously model the skeleton signals with long sequences. Experiments on several skeleton-based action classification benchmarks demonstrate the efficacy of our SUGAR. Moreover, experiments on zero-shot scenarios show that SUGAR is more versatile than linear-based methods.
Ensuring consistently high-quality training data is essential for developing reliable machine learning systems. Recent research demonstrates that incorporating human supervision into training set debugging effectively improves model performance, especially for text classification tasks. However, such methods often prove inapplicable to image understanding tasks, where inherently unstructured pixel data presents challenges in understanding and correcting biases. Inspired by human-AI alignment, we introduce AACA (Attribution Analysis-based Concept Alignment), a human-in-the-loop framework that mitigates bias in the training set by aligning the concepts used by humans and AI during the decision-making process. Specifically, AACA comprises two primary stages: interpretable data bug discovery and targeted data augmentation. During the data bug discovery stage, AACA identifies confounded and valid concepts to explain why prediction failure occurs and what concept the model should focus, using interpretability methods and human annotation. In the stage of targeted data augmentation, AACA adopts these concept-level attributions as clues to synthesize debugging instances via text-to-image generative model. The initial model is then retrained on the augmented set to correct prediction failures. Comparative experiments conducted on crowdsourced annotations and real-world datasets demonstrate that AACA can accurately identifies data bugs and effectively repairs prediction failures, thereby significantly improving prediction performance.
Explainable AI (XAI) methods generally fall into two categories. Post-hoc approaches generate explanations for pre-trained models and are compatible with various neural network architectures. These methods often use feature importance visualizations, such as saliency maps, to indicate which input regions influenced the model’s prediction. Unfortunately, they typically offer a coarse understanding of the model’s decision-making process. In contrast, ante-hoc (inherently explainable) methods rely on specially designed model architectures trained from scratch. A notable subclass of these methods provides explanations through prototypes, representative patches extracted from the training data. However, prototype-based approaches require dedicated architectures, involve specialized training procedures, and perform well only on specific datasets. In this work, we propose EPIC (Explanation of Pretrained Image Classification), a novel approach that bridges the gap between these two paradigms. Like post-hoc methods, EPIC operates on pre-trained models without architectural modifications. Simultaneously, it delivers intuitive, prototype-based explanations inspired by ante-hoc techniques. To the best of our knowledge, EPIC is the first post-hoc method capable of fully replicating the core explanatory power of inherently interpretable models. We evaluate EPIC on benchmark datasets commonly used in prototype-based explanations, such as CUB-200-2011 and Stanford Cars, alongside large-scale datasets like ImageNet, typically employed by post-hoc methods. EPIC uses prototypes to explain model decisions, providing a flexible and easy-to-understand tool for creating clear, high-quality explanations.
Facial Dynamics in Video: Instruction Tuning for Improved Facial Expression Perception and Contextual Awareness
PDF ↗Facial expression captioning has found widespread application across various domains. Recently, the emergence of video Multimodal Large Language Models (MLLMs) has shown promise in general video understanding tasks. However, describing facial expressions within videos poses two major challenges for these models: (1) the lack of adequate datasets and benchmarks, and (2) the limited visual token capacity of video MLLMs. To address these issues, this paper introduces a new instruction-following dataset tailored for dynamic facial expression caption. The dataset comprises 5,033 high-quality video clips annotated manually, containing over 700,000 tokens. Its purpose is to improve the capability of video MLLMs to discern subtle facial nuances. Furthermore, we propose FaceTrack-MM, which leverages a limited number of tokens to encode the main character’s face. This model demonstrates superior performance in tracking faces and focusing on the facial expressions of the main characters, even in intricate multiperson scenarios. Additionally, we introduce a novel evaluation metric combining event extraction, relation classification, and the longest common subsequence (LCS) algorithm to assess the content consistency and temporal sequence consistency of generated text. Moreover, we present FECBench, a benchmark designed to assess the performance of existing video MLLMs in this specific task.
Concept Bottleneck Models (CBMs) enhance interpretability by introducing a layer of human-understandable concepts between inputs and predictions. While recent methods automate concept generation using Large Language Models (LLMs) and Vision-Language Models (VLMs), they still face three fundamental challenges: poor visual grounding, concept redundancy, and the absence of principled metrics to balance predictive accuracy and concept compactness. We introduce PS-CBM, a Partially Shared CBM framework that addresses these limitations through three core components: (1) a multimodal concept generator that integrates LLM-derived semantics with exemplar-based visual cues; (2) a Partially Shared Concept Strategy that merges concepts based on activation patterns to balance specificity and compactness; and (3) Concept-Efficient Accuracy (CEA), a post-hoc metric that jointly captures both predictive accuracy and concept compactness. Extensive experiments on eleven diverse datasets show that PS-CBM consistently outperforms state-of-the-art CBMs, improving classification accuracy by 1.0%–7.4% and CEA by 2.0%–9.5%, while requiring significantly fewer concepts. These results underscore PS-CBM’s effectiveness in achieving both high accuracy and strong interpretability.
FaNe: Towards Fine-Grained Cross-Modal Contrast with False-Negative Reduction and Text-Conditioned Sparse Attention
PDF ↗Medical vision-language pre-training (VLP) offers significant potential for advancing medical image understanding by leveraging paired image-report data. However, existing methods are limited by False Negatives (FaNe) induced by semantically similar texts and insufficient fine-grained cross-modal alignment. To address these limitations, we propose FaNe, a semantic-enhanced VLP framework. To mitigate false negatives, we introduce a semantic-aware positive pair mining strategy based on text-text similarity with adaptive normalization. Furthermore, we design a text-conditioned sparse attention pooling module to enable fine-grained image-text alignment through localized visual representations guided by textual cues. To strengthen intra-modal discrimination, we develop a hard-negative aware contrastive loss that adaptively reweights semantically similar negatives. Extensive experiments on five downstream medical imaging benchmarks demonstrate that FaNe achieves state-of-the-art performance across image classification, object detection, and semantic segmentation, validating the effectiveness of our framework.