LLMs have demonstrated impressive performance across various language tasks. However, the strengths of LLMs can vary due to different architectures, model sizes, areas of training data, etc. Therefore, ensemble reasoning for the strengths of different LLM experts is critical to achieving consistent and satisfactory performance on diverse inputs across a wide range of tasks. However, existing LLM ensemble methods are either computationally intensive or incapable of leveraging complementary knowledge among LLM experts for various inputs. In this paper, we propose an efficient Dynamic Ensemble Reasoning paradigm, called DER to integrate the strengths of multiple LLM experts conditioned on dynamic inputs. Specifically, we model the LLM ensemble reasoning problem as a Markov Decision Process, wherein an agent sequentially takes inputs to request knowledge from an LLM candidate and passes the output to a subsequent LLM candidate. Moreover, we devise a reward function to train a DER-Agent to dynamically select an optimal answering route given the input questions, aiming to achieve the highest performance with as few computational resources as possible. Last, to fully transfer the expert knowledge from the prior LLMs, we develop a Knowledge Transfer Prompt that enables the subsequent LLM candidates to transfer complementary knowledge effectively. Experiments demonstrate that our method uses fewer computational resources to achieve better performance compared to state-of-the-art baselines. Code and appendix are available at https://github.com/Fhujinwu/DER.
论文检索
输入标题、作者或关键词，从 1,280 篇学术成果中精准定位
Efficient key-value (KV) cache compression is critical for scaling transformer-based Large Language Models (LLMs) in long sequences and resource-limited settings. Existing methods evict tokens based on their positions or importance, but position-based strategies can miss crucial information outside predefined regions, while those relying on global importance scores resulting in strong regional biases, limiting the KV cache's overall context retention and potentially impairing the performance of LLMs on complex tasks. Our wavelet analysis reveals that as tokens approach the end of sequence, their contributions to generation gradually increase and tends to diverge more from neighboring tokens, indicating a smooth transition with increasing complexity and variability from distant to nearby context. Motivated by this observation, we propose TreeKV, an intuitive, training-free method that employs a tree structure for smooth cache compression. TreeKV maintains a fixed cache size, allowing LLMs to deliver high-quality output in long text scenarios and is applicable during both the generation and prefilling stages. TreeKV consistently surpasses all baseline models in language modeling tasks on PG19 and OpenWebText2, allowing LLMs trained with short context window to generalize to longer window with a 16x cache reduction. On the Longbench benchmark, TreeKV achieves the best performance with only 6% of the budget at optimal efficiency.
Improving Prediction Certainty Estimation for Reliable Early Exiting via Null Space Projection
PDF ↗Early exiting has demonstrated great potential in accelerating the inference of pre-trained language models (PLMs) by enabling easy samples to exit at shallow layers, eliminating the need for executing deeper layers. However, existing early exiting methods primarily rely on class-relevant logits to formulate their exiting signals for estimating prediction certainty, neglecting the detrimental influence of class-irrelevant information in the features on prediction certainty. This leads to an overestimation of prediction certainty, causing premature exiting of samples with incorrect early predictions. To remedy this, we define an NSP score to estimate prediction certainty by considering the proportion of class-irrelevant information in the features. On this basis, we propose a novel early exiting method based on the Certainty-Aware Probability (CAP) score, which integrates insights from both logits and the NSP score to enhance prediction certainty estimation, thus enabling more reliable exiting decisions. The experimental results on the GLUE benchmark show that our method can achieve an average speed-up ratio of 2.19× across all tasks with negligible performance degradation, surpassing the state-of-the-art (SOTA) ConsistentEE by 28%, yielding a better trade-off between task performance and inference efficiency. The code is available at https://github.com/He-Jianing/NSP.git.
Bridging Generative and Discriminative Learning: Few-Shot Relation Extraction via Two-Stage Knowledge-Guided Pre-training
PDF ↗Few-Shot Relation Extraction (FSRE) remains a challenging task due to the scarcity of annotated data and the limited generalization capabilities of existing models. Although large language models (LLMs) have shown potential in FSRE through in-context learning, their general-purpose training objectives often result in suboptimal performance for task-specific relation extraction. To overcome these challenges, we propose TKRE (Two-Stage Knowledge-Guided Pre-training for Relation Extraction), a novel framework that synergistically integrates LLMs with traditional relation extraction models, bridging generative and discriminative learning paradigms. TKRE introduces two key innovations: (1) leveraging LLMs to generate explanation-driven knowledge and schema-constrained synthetic data, addressing the issue of data scarcity; and (2) a two-stage pre-training strategy combining Masked Span Language Modeling (MSLM) and Span-Level Contrastive Learning (SCL) to enhance relational reasoning and generalization. Together, these components enable TKRE to effectively handle FSRE tasks. Comprehensive experiments on benchmark datasets demonstrate the efficacy of TKRE, achieving new state-of-the-art performance in FSRE and underscoring its potential for broader application in low-resource scenarios. The code and data are released on https://github.com/UESTC-GQJ/TKRE.
Large Language Models (LLMs) have achieved remarkable success across various domains. However, they still face significant challenges, including high computational costs for training and limitations in solving complex reasoning problems. Although existing methods have extended the reasoning capabilities of LLMs through structured paradigms, these approaches often rely on task-specific prompts and predefined reasoning processes, which constrain their flexibility and generalizability. To address these limitations, we propose a novel framework that leverages graph learning to enable more flexible and adaptive reasoning capabilities for LLMs. Specifically, this approach models the reasoning process of a problem as a graph and employs LLM-based graph learning to guide the adaptive generation of each reasoning step. To further enhance the adaptability of the model, we introduce a Graph Neural Network (GNN) module to perform representation learning on the generated reasoning process, enabling real-time adjustments to both the model and the prompt. Experimental results demonstrate that this method significantly improves reasoning performance across multiple tasks without requiring additional training or task-specific prompt design. Code can be found in https://github.com/zch65458525/L2T.
SecV: LLM-based Secure Verilog Generation with Clue-Guided Exploration on Hardware-CWE Knowledge Graph
PDF ↗Verilog is specified as the primary Register Transfer Level (RTL) hardware description language, which designs the logical functions between registers for digital circuit systems. Recently, there emerges much cutting-edge research in leveraging Large Language Models (LLMs) to generate Verilog, aiming at effectively reducing errors and costs in the logic design of chips. However, these works mainly focus on logical correctness or PPA (Power, Performance, Area) measurement of the generated results, while neglecting the security problems in Verilog. In this study, we propose SecV, a novel and unified framework to generate secure Verilog by clue-guided exploration on Common Weakness Enumeration (CWE) knowledge graph (KG) for chips. First, the builder of the KG utilizes the instance-adapted chain of thought (COT) to extract entities and their relationships from raw Hardware-CWE corpora. Then, a fine-tuned BERT model is employed to verify the Hardware-CWE KG and collaborate with builder iteratively to achieve the precise KG. Based on Hardware-CWE KG, a clue-guided graph exploration paradigm is designed to facilitate collaborative inference of knowledge to generate secure Verilog by LLMs. Experiments demonstrate that SecV achieves 82.6% secure Verilog code without specified CWE in the generated functionally correct Verilog, with superior performance of a 21.7% performance improvement compared to SOTA.
M3ANet: Multi-scale and Multi-Modal Alignment Network for Brain-Assisted Target Speaker Extraction
PDF ↗The brain-assisted target speaker extraction (TSE) aims to extract the attended speech from mixed speech by utilizing the brain neural activities, for example Electroencephalography (EEG). However, existing models overlook the issue of temporal misalignment between speech and EEG modalities, which hampers TSE performance. In addition, the speech encoder in current models typically uses basic temporal operations (e.g., one-dimensional convolution), which are unable to effectively extract target speaker information. To address these issues, this paper proposes a multi-scale and multi-modal alignment network (M3ANet) for brain-assisted TSE. Specifically, to eliminate the temporal inconsistency between EEG and speech modalities, the modal alignment module that uses a contrastive learning strategy is applied to align the temporal features of both modalities. Additionally, to fully extract speech information, multi-scale convolutions with GroupMamba modules are used as the speech encoder, which scans speech features at each scale from different directions, enabling the model to capture deep sequence information. Experimental results on three publicly available datasets show that the proposed model outperforms current state-of-the-art methods across various evaluation metrics, highlighting the effectiveness of our proposed method. The source code is available at: https://github.com/fchest/M3ANet.
Temporal claims, often riddled with inaccuracies, are a significant challenge in the digital misinformation landscape. Fact-checking systems that can accurately verify such claims are crucial for combating misinformation. Current systems struggle with the complexities of evaluating the accuracy of these claims, especially when they include multiple, overlapping, or recurring events. We introduce a novel timeline-based fact verification framework that identify events from both claim and evidence and organize them into their respective chronological timelines. The framework systematically examines the relationships between the events in both claim and evidence to predict the veracity of each claim event and their chronological accuracy. This allows us to accurately determine the overall veracity of the claim. We also introduce a new dataset of complex temporal claims involving timeline-based reasoning for the training and evaluation of our proposed framework. Experimental results demonstrate the effectiveness of our approach in handling the intricacies of temporal claim verification.
Scalable Multi-Stage Influence Function for Large Language Models via Eigenvalue-Corrected Kronecker-Factored Parameterization
PDF ↗Pre-trained large language models (LLMs) are commonly fine-tuned to adapt to downstream tasks. Since the majority of knowledge is acquired during pre-training, attributing the predictions of fine-tuned LLMs to their pre-training data may provide valuable insights. Influence functions have been proposed as a means to explain model predictions based on training data. However, existing approaches often fail to compute "multi-stage" influence and lack scalability to billion-scale LLMs. In this paper, we propose multi-stage influence functions to attribute the downstream predictions of fine-tuned LLMs to pre-training data under the full-parameter fine-tuning paradigm. To enhance the efficiency and practicality of our multi-stage influence function, we leverage Eigenvalue-corrected Kronecker-Factored (EK-FAC) parameterization for efficient approximation. Empirical results validate the superior scalability of EK-FAC approximation and the effectiveness of our multi-stage influence function. Additionally, case studies on a real-world LLM, dolly-v2-3b, demonstrate its interpretive power, with exemplars illustrating insights provided by multi-stage influence estimates.
Multimodal news contains a wealth of information and is easily affected by deepfake modeling attacks. To combat the latest image and text generation methods, we present a new Multimodal Fake News Detection dataset (MFND) containing 11 manipulated types, designed to detect and localize highly authentic fake news. Furthermore, we propose a Shallow-Deep Multitask Learning (SDML) model for fake news, which fully uses unimodal and mutual modal features to mine the intrinsic semantics of news. Under shallow inference, we propose the momentum distillation-based light punishment contrastive learning for fine-grained uniform spatial image and text semantic alignment, and an adaptive cross-modal fusion module to enhance mutual modal features. Under deep inference, we design a two-branch framework to augment the image and text unimodal features, respectively merging with mutual modalities features, for four predictions via dedicated detection and localization projections. Experiments on both mainstream and our proposed datasets demonstrate the superiority of the model. Codes and dataset are released at https://github.com/yunan-wang33/sdml.
Arrhythmia diagnosis using electrocardiogram (ECG) is critical for preventing cardiovascular risks. However, existing deep learning-based methods struggle with label scarcity and contrastive learning-based methods suffer from false-negative samples, which lead to poor model generalization. Besides, due to inter-subject variability, pre-trained models cannot achieve evenly performance across individuals. Conducting model fine-tuning for each individual is computationally expensive and does not guarantee improvement. We propose DiffECG, a diffusion-based self-supervised learning framework for label-efficient and personalized arrhythmia detection. Our method utilizes a diffusion model to extract robust ECG representations, coupled with a novel feature extractor and a multi-modal feature fusion strategy to obtain a well-generalized model. Moreover, we propose an efficient model personalization mechanism based on zeroth-order optimization. It personalizes the model by tuning the noise-adding step t in the diffusion process, significantly reducing computational costs compared to model fine-tuning. Experimental results show that our proposed method outperforms the SOTA method by 37.9% and 23.9% in generalization and personalization performance, respectively. The source code is available at: https://github.com/Auguuust/DiffEC
Dense object segmentation is essential for various applications, particularly in pathology image and remote sensing image analysis. However, distinguishing numerous similar and densely packed objects in this task presents significant challenges. Several methods, including CNN- and ViT-based approaches, have been proposed to tackle these issues. Yet, models trained on limited datasets exhibit limited generalization ability. The Segment Anything Model (SAM) has recently achieved significant progress in zero-shot segmentation but relies heavily on precise positional guidance. However, providing numerous accurate location prompts in dense scenarios is time-consuming. To overcome this limitation, we conducted an in-depth exploration of the SAM mechanism and found that its strong generalization ability stems from the encoder’s edge detection capability, which is semantically independent, making location prompts essential for segmentation. This insight inspired the development of DenseSAM, which replaces location prompts with semantic guidance for automatic segmentation in dense scenarios. Specifically, it uses local details to weaken the edges of background objects, leverages global context to enhance intra-class feature similarity, while further increasing contrast with the background, and integrates a dual-head decoding process to enable lightweight automatic semantic segmentation. Extensive experiments on pathology images demonstrate that DenseSAM delivers remarkable performance with minimal training parameters, providing a cost-effective and efficient solution. Moreover, experiments on remote sensing images further validate its excellent scalability, making DenseSAM suitable for various dense object segmentation domains. The code is available at https://github.com/imAzhou/DenseSAM.
Nighttime semantic segmentation is a critical yet challenging task in autonomous driving. Most existing methods are designed for daytime scenarios, resulting in poor nighttime performance due to texture loss and decreased object visibility. Low-light enhancement was applied before segmentation but failed to recover nighttime-specific details, introducing noise or losing delicate structures. Recent work shows that large-scale image-text pairs can effectively leverage natural language priors to guide visual representation, achieving remarkable performance across various downstream visual tasks. However, effectively employing visual-linguistic priors for nighttime semantic segmentation remains underexplored. To address these issues, we propose Text-WaveletFormer, a novel end-to-end framework that integrates text prompts and wavelet-based texture enhancement. Specifically, to compensate for the low recognizability of objects in nighttime scenes, we design a Text-Image Fusion Module (TIFM) to incorporate textual priors to improve nighttime object recognition. In addition, to alleviate the lack of texture details in nighttime conditions, we introduce a Wavelet Guided Texture Amplifier Module (WTAM) to fuse wavelet and raw image features via cross-attention, restoring low-light details. Finally, extensive experiments on benchmarks including NightCity, NightCity-fine, BDD100K, and CityScapes demonstrate our method’s superior performance over existing approaches.
Satisfiability Modulo Theories (SMT) solvers are crucial in many applications, yet their performance is often a bottleneck. This paper introduces SIRISMT, a novel framework that employs machine learning techniques for the automatic synthesis of efficient SMT-solving strategies. Specifically, SIRISMT targets at Z3 and consists of three key stages. First, given a set of training SMT formulas, SIRISMT simulates the solving process by leveraging reinforcement learning to guide its exploration within the strategy space. Next, SIRISMT refines the collected strategies by pruning redundant tactics and generating augmented strategies based on the subsequence structure of the learned strategies. These refined strategies are then fed back into the reinforcement learning model. Finally, the refined and optimized strategies are integrated into one strategy, which can be directly plugged into modern SMT solvers. Extensive evaluations show the superior performance of SIRISMT over the baseline methods. For example, compared to the default Z3, it solves 26.8% more formulas and achieves up to an 86.3% improvement in the Par-2 score on benchmark datasets. Additionally, we show that the synthesized strategy can improve the code coverage by up to 11.8% in a downstream symbolic execution benchmark.
MMGIA: Gradient Inversion Attack Against Multimodal Federated Learning via Intermodal Correlation
PDF ↗Multimodal federated learning (MMFL) enables collaborative model training across multiple modalities, such as images and text, without requiring direct data sharing. However, the inherent correlations between modalities introduce new privacy vulnerabilities, making MMFL more susceptible to gradient inversion attacks. In this work, we propose MMGIA, an intermodal correlation-driven gradient inversion attack that systematically exploits multimodal correlation to enhance data reconstruction quality. MMGIA consists of a two-stage optimization framework: the first stage independently reconstructs each modality using traditional gradient inversion techniques, while the second stage refines these reconstructions through pre-trained feature extractors to align modalities in a shared latent space. To further improve reconstruction accuracy, we introduce a quality-weighted fusion strategy, which dynamically integrates multimodal embeddings into a global fused representation that serves as a guiding signal for refining each modality’s reconstruction. This ensures that high-quality reconstructions contribute more to the optimization process, preventing degradation in well-reconstructed modalities while enhancing weaker ones. We conduct extensive experiments on multiple multimodal scenarios, demonstrating that MMGIA outperforms both the only existing multimodal attack and state-of-the-art single-modal attacks, revealing the heightened privacy risks in MMFL.
Although deep learning models have demonstrated remarkable potential in weather prediction, most of them overlook either the physics of the underlying weather evolution or the topology of the Earth’s surface. In light of these disadvantages, we develop PASSAT, a novel Physics-ASSisted And Topology-informed deep learning model for weather prediction. PASSAT attributes the weather evolution to two key factors: (i) the advection process that can be characterized by the advection equation and the Navier-Stokes equation; (ii) the Earth-atmosphere interaction that is difficult to both model and calculate. PASSAT also takes the topology of the Earth's surface into consideration, other than simply treating it as a plane. With these considerations, PASSAT numerically solves the advection equation and the Navier-Stokes equation on the spherical manifold, utilizes a spherical graph neural network to capture the Earth-atmosphere interaction, and generates the initial velocity fields that are critical to solving the advection equation from the same spherical graph neural network. In the 5.625-degree resolution ERA5 data set, PASSAT outperforms both the state-of-the-art deep learning-based weather prediction models and the operational numerical weather prediction model IFS T42.
Graph Neural Networks (GNNs) are known to be prone to adversarial attacks, among which backdoor attack is a major security threat. By injecting backdoor triggers into a graph and assigning a target class label to nodes attached to the triggers, the attacker can mislead the GNN model trained on the poisoned graph to classify test nodes attached with a trigger to the target class. To defend against backdoor attacks, existing defense methods rely on anomaly detection in feature distribution or label transformation. However, these approaches are incapable of detecting in-distribution triggers or clean-label attacks that do not alter the class label of target nodes. To tackle these threats, we empirically analyze triggers from a multidimensional aspect, and our analysis shows that there are clear distinctions between trigger nodes and normal ones in terms of node feature values, node embeddings, and class prediction probabilities. Based on these findings, we propose a Multidimensional Anomaly Detection framework (MAD) that can effectively minimize the impact of triggers by pruning away anomalous nodes and edges. Extensive experiments show that at the cost of slight loss in clean classification accuracy, MAD achieves considerably lower attack success rate as compared to state-of-the-art backdoor defense methods.
Multimodal Inverse Attention Network with Intrinsic Discriminant Feature Exploitation for Fake News Detection
PDF ↗Multimodal fake news detection has garnered significant attention due to its profound implications for social security. While existing approaches have contributed to understanding cross-modal consistency, they often fail to leverage modal-specific representations and explicit discrepant features. To address these limitations, we propose a Multimodal Inverse Attention Network (MIAN), a novel framework that explores intrinsic discriminative features based on news content to advance fake news detection. Specifically, MIAN introduces a hierarchical learning module that captures diverse intra-modal relationships through local-to-global and local-to-local interactions, thereby generating enhanced unimodal representations to improve the identification of fake news at the intra-modal level. Additionally, a cross-modal interaction module employs a co-attention mechanism to establish and model dependencies between the refined unimodal representations, facilitating seamless semantic integration across modalities. To explicitly extract inconsistency features, we propose an inverse attention mechanism that effectively highlights the conflicting patterns and semantic deviations introduced by fake news in both intra- and inter-modality. Extensive experiments on benchmark datasets demonstrate that MIAN significantly outperforms state-of-the-art methods, underscoring its pivotal contribution to advancing social security through enhanced multimodal fake news detection.
Subgraph Information Bottleneck with Causal Dependency for Stable Molecular Relational Learning
PDF ↗Molecular Relational Learning (MRL) is widely applied in molecular sciences. Recent studies attempt to retain molecular core information (e.g., substructures) by Graph Information Bottleneck but primarily focus on information compression without considering the causal dependencies of chemical reactions among substructures. This oversight neglects the core factors that determine molecular relationships, making maintaining stable MRL in distribution-shifted data challenging. To bridge this gap, we propose the Causal Subgraph Information Bottleneck (CausalGIB) for stable MRL. CausalGIB leverages causal dependency to guide substructure representation and integrates subgraph information bottleneck to optimize the core substructure representation, generating stable representations. Specifically, we distinguish causal and confounding substructures by noise injection and substructure interaction based on causal analysis. Furthermore, by minimizing the discrepancy between causal and confounding information within subgraph information bottleneck, CausalGIB captures core substructures composed of causal substructures and aggregates them into molecular representations to improve their stability. Experimental results on nine datasets demonstrate that CausalGIB outperforms state-of-the-art models in two tasks and significantly enhances model’s stability in distribution-shifted data.
Reconstructing visual stimuli from EEG signals is a crucial step in realizing brain-computer interfaces. In this paper, we propose a transformer-based EEG signal encoder integrating the Discrete Wavelet Transform (DWT) and the gating mechanism. Guided by the feature alignment and category-aware fusion losses, this encoder is used to extract features related to visual stimuli from EEG signals. Subsequently, with the aid of a pre-trained diffusion model, these features are reconstructed into visual stimuli. To verify the effectiveness of the model, we conducted EEG-to-image generation and classification tasks using the THINGS-EEG dataset. To address the limitations of quantitative analysis at the semantic level, we combined WordNet-based classification and semantic similarity metrics to propose a novel semantic-based score, emphasizing the ability of our model to transfer neural activities into visual representations. Experimental results show that our model significantly improves semantic alignment and classification accuracy, which achieves a maximum single-subject accuracy of 43%, outperforming other state-of-the-art methods. The source code is available at https://github.com/zes0v0inn/DWT_EEG_Reconstruction/.