Retrieval-Augmented Generation (RAG) has become a key paradigm for reducing factual hallucinations in Large Language Models (LLMs), yet little is known about how the order of retrieved documents affects model behavior.We empirically show that under a Top-5 retrieval setting with the gold document included, LLM answers vary substantially across permutations of the retrieved set, even when the gold document is fixed in the first position.This reveals a previously underexplored sensitivity to retrieval permutations.Although existing robust RAG methods focus primarily on enhancing LLM robustness to low-quality retrieval and mitigating positional bias to distribute attention fairly over long contexts, neither approach directly addresses permutation sensitivity.In this paper, we propose Stable-RAG, which exploits permutation sensitivity estimation to mitigate permutation-induced hallucinations.Stable-RAG runs the generator under multiple retrieval orders, clusters hidden states, and decodes from a cluster-center representation that captures the dominant reasoning pattern.It then uses these reasoning results to align hallucinated outputs toward the correct answer, encouraging the model to produce consistent and accurate predictions across document permutations.Experiments on three QA datasets show that Stable-RAG improves answer accuracy, reasoning consistency, and generalization across datasets, retrievers, and input lengths compared with strong baselines 1 .

Efficiently editing the knowledge of large language models (LLMs) is crucial for real-world adaptation and error correction, yet remains a fundamental challenge. Existing approaches, primarily based on the Locate-Then-Edit paradigm, have achieved promising results by first identifying key parameters and then updating them; however, they still suffer from catastrophic over-editing, where sequential updates degrade both preserved and edited knowledge, especially as batch sizes grow. To address this, we propose Important Weight Constrained Knowledge Editing (IWCKEdit), a simple yet effective framework that introduces weight protection into the model editing process. IWCKEdit comprises two key modules: a front-end Protected Weight Identification Module that leverages Fisher information to detect weights essential for knowledge retention and effective editing, and a back-end Constrained Weight Updating Module that safeguards these critical weights during targeted updates for knowledge editing. Extensive experiments on large-scale models including GPT2-XL, Llama3-8B, and Qwen2.5-7B demonstrate that IWCKEdit achieves superior knowledge editing accuracy while robustly preserving both newly introduced and existing knowledge, substantially advancing the state of the art in LLM editing.

Current RAG retrievers are designed primarily for human readers, emphasizing complete, readable, and coherent paragraphs. However, Large Language Models (LLMs) benefit more from precise, compact, and well-structured input, which enhances reasoning quality and efficiency. Existing methods rely on reranking or summarization to identify key sentences, but may introduce semantic breaks and unfaithfulness. Thus, efficiently extracting and organizing answer-relevant clues from large-scale documents while reducing LLM reasoning costs remains challenging in RAG systems. Inspired by Occam's razor, we frame LLM-centric retrieval as MinMax optimization: maximizing the extraction of potential clues and reranking them for well-organization, while minimizing reasoning costs by truncating to the smallest sufficient set of clues. In this paper, we propose CompSelect, a compact clue selection mechanism for LLM-centric RAG, consisting of a clue extractor, a reranker, and a truncator. (1) The clue extractor first uses answer-containing sentences as fine-tuning targets, aiming to extract sufficient potential clues; (2) The reranker is trained to prioritize effective clues based on real LLM feedback; (3) The truncator uses the truncated text containing the minimum sufficient clues for answering the question as fine-tuning targets, thereby enabling efficient RAG reasoning. Experiments on three QA datasets demonstrate that CompSelect improves performance while reducing both total and online latency compared to a range of baseline methods. Further analysis also confirms its robustness to unreliable retrieval and generalization across different scenarios.

In affective neuroscience and emotion-aware AI, understanding how complex auditory stimuli drive emotion arousal dynamics remains unresolved. This study introduces a neurobiologically informed computational framework to model the brain's encoding of naturalistic auditory inputs into dynamic behavioral/neural responses using four datasets (SEED, LIRIS, self-collected SEED Annotation and BAVE). Guided by neurobiological principles of parallel auditory hierarchy, we decompose audio into multilevel auditory features (through classical algorithms and wav2vec 2.0/Hubert) from the original and isolated human voice/background soundtrack elements, mapping them to emotion-related responses via cross-dataset analyses. Our analysis reveals that high-level semantic representations (derived from the final layer of wav2vec 2.0/Hubert) exert a dominant role in emotion encoding, outperforming low-level acoustic features with significantly stronger mappings to behavioral annotations and dynamic neural synchrony across most brain regions ($p \lt 0.05$). Notably, middle layers of wav2vec 2.0/hubert (balancing acoustic-semantic information) surpass the final layers in emotion induction across datasets. Moreover, human voices and soundtracks show dataset-dependent emotion-evoking biases aligned with stimulus energy distribution (e.g., LIRIS favors soundtracks due to higher background energy), with neural analyses indicating voices dominate prefrontal/temporal activity while soundtracks excel in limbic regions. By integrating affective computing and neuroscience, this work uncovers hierarchical mechanisms of auditory-emotion encoding, providing a foundation for adaptive emotion-aware systems and cross-disciplinary explorations of audio-affective interactions.

In a real-world RAG system, the current query often involves spoken ellipses and ambiguous references from dialogue contexts, necessitating query rewriting to better describe user's information needs. However, traditional context-based rewriting has minimal enhancement on downstream generation tasks due to the lengthy process from query rewriting to response generation. Some researchers try to utilize reinforcement learning with generation feedback to assist the rewriter, but this sparse rewards provide little guidance in most cases, leading to unstable training and generation results.We find that user's needs are also reflected in the gold documents, retrieved documents and ground-truth. Therefore, by feeding back these multi-aspect dense rewards to query rewriting, more stable and satisfactory responses can be achieved. In this paper, we propose a novel query rewriting method MaFeRw, which improves RAG performance by integrating multi-aspect feedback from both the retrieval process and generated results. Specifically, we first use manual data to train a T5 model for the rewriter initialization. Next, we design three metrics as reinforcement learning feedback: the similarity between the rewritten query and the gold document, the ranking metrics, and ROUGE between the generation and the ground truth. Inspired by RLAIF, we train three kinds of reward models for the above metrics to achieve more efficient training. Finally, we combine the scores of these reward models as feedback, and use PPO algorithm to explore the optimal query rewriting strategy.Experimental results on two conversational RAG datasets demonstrate that MaFeRw achieves superior generation metrics and more stable training compared to baselines.

The scaled Web 3.0 digital economy, represented by decentralized finance (DeFi), has sparked increasing interest in the past few years, which usually relies on blockchain for token transfer and diverse transaction logic. However, illegal behaviors, such as financial fraud, hacker attacks, and money laundering, are rampant in the blockchain ecosystem and seriously threaten its integrity and security. In this paper, we propose a novel double graph-based Ethereum account de-anonymization inference method, dubbed DBG4ETH, which aims to capture the behavioral patterns of accounts comprehensively and has more robust analytical and judgment capabilities for current complex and continuously generated transaction behaviors. Specifically, we first construct a global static graph to build complex interactions between the various account nodes for all transaction data. Then, we also construct a local dynamic graph to learn about the gradual evolution of transactions over different periods. Different graphs focus on information from different perspectives, and features of global and local, static and dynamic transaction graphs are available through DBG4ETH. In addition, we propose an adaptive confidence calibration method to predict the results by feeding the calibrated weighted prediction values into the classifier. Experimental results show that DBG4ETH achieves state-of-the-art results in the account identification task, improving the F1-score by at least 3.75% and up to 40.52% compared to processing each graph type individually and outperforming similar account identity inference methods by 5.23 % to 12.91 %.

With the flourishing of the Agricultural Internet of Things (AIoT), analyzing large-volume sensor data has become a regular requirement for agricultural decision-making. Federated learning (FL), which facilitates scattered AIoT devices to train models collaboratively, has gained significant attention. However, traditional FL poses challenges in AIoT scenarios, such as wide geo-distribution, heterogeneous data distribution, and high-device risks. Existing works tend to be one-sided and remain unclear on how to tackle these issues thoroughly in AIoT. To fill the gap, we present Bc2FL, a double-layer blockchain-based FL framework, which enhances both learning efficiency and security for AIoT. The double-layer blockchain, coupled with a two-stage consensus algorithm, drives the hierarchical FL process to enable efficient and reliable agricultural knowledge-sharing. In addition, Bc2FL adopts an adaptive model aggregation algorithm to dynamically tune noise levels based on the model quality, further improving the learning security and model credibility. Finally, the extensive experimental results demonstrate that Bc2FL not only improves the model accuracy by up to 21.17% compared with the state-of-the-art baselines, but also enhances the privacy protection within an additional error of only 2.1%.

In the Web 3.0 ecosystem, blockchain and Artificial Intelligence of Things (AIoT) construct the infrastructure, where blockchain transaction semantic detection (BTSD) aims to enhance blockchain security by identifying illegal transactions through distributed miners executing AI algorithms. However, the computational cost of performing semantic detection discourages miners from participating without adequate incentives. Existing studies focus on algorithmic aspects of BTSD, which generally ignore the critical issue of incentive mechanism. To fill this gap, we propose the first incentive-based BTSD framework in the transaction pool phase, emphasizing how incentives affect the behavior of miners and users. We use evolutionary game theory to model miner-user interactions and define three key scenarios to simulate the impact of reward decay and penalty factors on system dynamics. Our results demonstrate that adjusting these parameters significantly influences the number of miners engaging in semantic detection and users initiating legitimate transactions. Under certain conditions, a well-designed incentive mechanism can lead to an Evolutionary Stable Strategy (ESS), thereby achieving systemic stability. This study introduces a novel incentive mechanism for BTSD during the transaction pool phase and validates its effectiveness through both theoretical insights and numerical solutions to enhance blockchain security.

Private data, being larger and quality-higher than public data, can greatly improve large language models (LLM).However, due to privacy concerns, this data is often dispersed in multiple silos, making its secure utilization for LLM training a challenge.Federated learning (FL) is an ideal solution for training models with distributed private data, but traditional frameworks like FedAvg are unsuitable for LLM due to their high computational demands on clients.An alternative, split learning, offloads most training parameters to the server while training embedding and output layers locally, making it more suitable for LLM.Nonetheless, it faces significant challenges in security and efficiency.Firstly, the gradients of embeddings are prone to attacks, leading to potential reverse engineering of private data.Furthermore, the server's limitation of handle only one client's training request at a time hinders parallel training, severely impacting training efficiency.In this paper, we propose a Federated Learning framework for LLM, named FL-GLM, which prevents data leakage caused by both serverside and peer-client attacks while improving training efficiency.Specifically, we first place the input block and output block on local client to prevent embedding gradient attacks from server.Secondly, we employ key-encryption during client-server communication to prevent reverse engineering attacks from peer-clients.Lastly, we employ optimization methods like client-batching or server-hierarchical, adopting different acceleration methods based on the actual computational capabilities of the server.Experimental results on NLU and generation tasks demonstrate that FL-GLM achieves comparable metrics to centralized chatGLM model, validating the effectiveness of our federated learning framework.

How to improve accuracy of location prediction and select optimal prediction anchor, which has been a great challenge in dense object detection studies and applications. Nowadays, most object detection algorithms obtain bboxes location prediction (x1,y1, x2,y2) by multiple convolution of regression features. Through experiments in this paper, we demonstrate that x-bounds (x1, x2) and y-bounds (y1, y2) of the bboxes location regression results are affected by different features with different significance. Based on our studies, we propose a new detector to decouple the regression features to obtain feat_x and feat_y, which are used to predict the x-bounds and y-bounds of the bboxes, respectively. Besides, our experiments also show that the point with the most accurate x-bounds prediction is often not the point with the most accurate y-bounds prediction. For this reason, we have incorporated two additional branches to assess the prediction quality of x-bounds and y-bounds respectively, while designing the novel target functions and the loss functions for each branch. In contrast to the traditional methods of all four bounds predicted by a single point, this paper combine the resulting x-bounds and y-bounds by self-invented combination strategy, and the optimal box location prediction can be obtained. The method proposed in this paper achieved an accuracy of 48.2% with backbone "ResNeXt-101-64×4d" in 24 epochs training, which is a 0.6% improvement compared to TOOD.

As the coverage of permissioned blockchains expands and the number of participating replicas increases, a scalable and efficient Byzantine Fault Tolerant (BFT) protocol is essential for large-scale blockchain. Unfortunately, previous BFT consensus protocols rely on a single leader to drive the protocol, which becomes a bottleneck for system scalability when the number of replicas exceeds a certain threshold. Although some proposals suggest hierarchically grouping nodes into different layers to alleviate verification pressure on the single leader. However, existing solutions only support serial execution between layers, causing performance and latency bottlenecks. To address these issues, we propose TierFlow, the first layered consensus protocol that supports pipelined execution, maintaining high throughput in scenarios with a large-scale deployment of replicas. TierFlow innovatively addresses the serial execution bottleneck in layered consensus by decoupling inter-layer consensus. To eliminate redundant phases, we use a pre-proof method to advance the next round of verification, and utilize delayed verification to merge similar verification workflows. We implement TierFlow and compare it with advanced BFT protocols such as HotStuff and Fast-HotStuff. We conduct extensive experiments with over 100 replicas, demonstrating that TierFlow achieves throughput 14x higher than Fast-HotStuff in large-scale application scenarios, with the performance disparity widening as scale increases.

Temporal Knowledge Graph (TKG) reasoning involves predicting future facts based on historical information by learning correlations between entities and relations. Recently, many models have been proposed for the TKG reasoning task. However, most existing models cannot efficiently utilize historical information, which can be summarized in two aspects: 1) Many models only consider the historical information in a fixed time range, resulting in a lack of useful information; 2) some models use all the historical facts, thus some noise or invalid facts are introduced during reasoning. In this regard, we propose a novel TKG reasoning model with dynamic memory enhancement (DyMemR). Inspired by human memory, we introduce memory capacity, memory loss, and repetition stimulation to design a human-like memory pool that could remember potentially useful historical facts. To fully leverage the memory pool, we utilize a two-stage training strategy.The first stage is guided by the memory-based encoding module which learns embeddings from memory-based subgraphs generated through the memory pool. The second stage is the memory-based scoring module that emphasizes the historical facts in the memory pool. Finally, we extensively validate the superiority of DyMemR against various state-of-the-art baselines.

Embedding-based methods currently achieved impressive success in recommender systems. However, such methods are more likely to suffer from bias in data distribution, especially the attribute bias problem. For example, when a certain type of user, like the elderly, occupies the mainstream, the recommendation results of minority users would be seriously affected by the mainstream users’ attributes. To address this problem, most existing methods are proposed from the perspective of fairness, which focuses on eliminating unfairness but deteriorates the recommendation performance. Unlike these methods, in this paper, we focus on improving the recommendation performance for minority users of biased attributes. Along this line, we propose a novel attribute-aware Counterfactual Augmentation framework for Minority Users(CAMUS). Specifically, the CAMUS consists of a counterfactual augmenter, a confidence estimator, and a recommender. The counterfactual augmenter conducts data augmentation for the minority group by utilizing the interactions of mainstream users based on a universal counterfactual assumption. Besides, a tri-training-based confidence estimator is applied to ensure the effectiveness of augmentation. Extensive experiments on three real-world datasets have demonstrated the superior performance of the proposed methods. Further case studies verify the universality of the proposed CAMUS framework on different data sparsity, attributes, and models.

Stimulation to local areas remarkably affects brain activity patterns, which can be exploited to investigate neural bases of cognitive function and modify pathological brain statuses. There has been growing interest in exploring the fundamental action mechanisms of local stimulation. Nevertheless, how noise amplitude, an essential element in neural dynamics, influences stimulation-induced brain states remains unknown. Here, we systematically examine the effects of local stimulation by using a large-scale biophysical model under different combinations of noise amplitudes and stimulation sites. We demonstrate that noise amplitude nonlinearly and heterogeneously tunes the stimulation effects from both regional and network perspectives. Furthermore, by incorporating the role of the anatomical network, we show that the peak frequencies of unstimulated areas at different stimulation sites averaged across noise amplitudes are highly positively related to structural connectivity. Crucially, the association between the overall changes in functional connectivity as well as the alterations in the constraints imposed by structural connectivity with the structural degree of stimulation sites is nonmonotonically influenced by the noise amplitude, with the association increasing in specific noise amplitude ranges. Moreover, the impacts of local stimulation of cognitive systems depend on the complex interplay between the noise amplitude and average structural degree. Overall, this work provides theoretical insights into how noise amplitude and network structure jointly modulate brain dynamics during stimulation and introduces possibilities for better predicting and controlling stimulation outcomes.