Ontology classification,the problem of computing the subsumption hierarchies for classes (atomic concepts),is a core reasoning service provided by Web Ontology Language (OWL)reasoners.Although general-purpose OWL 2 re...Ontology classification,the problem of computing the subsumption hierarchies for classes (atomic concepts),is a core reasoning service provided by Web Ontology Language (OWL)reasoners.Although general-purpose OWL 2 reasoners employ sophisticated optimizations for classification,they are still not efficient owing to the high complexity of tableau algorithms for expressive ontologies. Profile-specific OWL 2 EL reasoners are efficient;however, they become incomplete even if the ontology contains only a small number of axioms that are outside the OWL 2 EL fragment.In this paper,we present a technique that combines an OWL 2 EL reasoner with an OWL 2 reasoner for ontology classification of expressive SROIQ.To optimize the workload,we propose a task decomposition strategy for identifying the minimal non-EL subontology that contains only necessary axioms to ensure completeness.During the ontology classification,the bulk of the workload is delegated to an efficient OWL 2 EL reasoner and only the minimal non- EL subontology is handled by a less efficient OWL 2 reasoner.The proposed approach is implemented in a prototype ComR and experimental results show that our approach offers a substantial speedup in ontology classification.For the wellknown ontology NCI,the classification time is reduced by 96.9%(resp.83.7%)compared against the standard reasoner Pellet (resp.the modular reasoner MORe).展开更多
Multimodal dialogue systems often fail to maintain coherent reasoning over extended conversations and suffer from hallucination due to limited context modeling capabilities.Current approaches struggle with crossmodal ...Multimodal dialogue systems often fail to maintain coherent reasoning over extended conversations and suffer from hallucination due to limited context modeling capabilities.Current approaches struggle with crossmodal alignment,temporal consistency,and robust handling of noisy or incomplete inputs across multiple modalities.We propose Multi Agent-Chain of Thought(CoT),a novel multi-agent chain-of-thought reasoning framework where specialized agents for text,vision,and speech modalities collaboratively construct shared reasoning traces through inter-agent message passing and consensus voting mechanisms.Our architecture incorporates self-reflection modules,conflict resolution protocols,and dynamic rationale alignment to enhance consistency,factual accuracy,and user engagement.The framework employs a hierarchical attention mechanism with cross-modal fusion and implements adaptive reasoning depth based on dialogue complexity.Comprehensive evaluations on Situated Interactive Multi-Modal Conversations(SIMMC)2.0,VisDial v1.0,and newly introduced challenging scenarios demonstrate statistically significant improvements in grounding accuracy(p<0.01),chain-of-thought interpretability,and robustness to adversarial inputs compared to state-of-the-art monolithic transformer baselines and existing multi-agent approaches.展开更多
A fairly precise vision of a healthy physical self can serve as motivation for undertaking the means to that end.The same cannot be said with regard to“healthy moral selves”.By definition,democracy is about living w...A fairly precise vision of a healthy physical self can serve as motivation for undertaking the means to that end.The same cannot be said with regard to“healthy moral selves”.By definition,democracy is about living with others and,as we argue,a healthy moral self is one that lives well with others.However,precisely what“living with others”entails is ambiguous,particularly in a capitalist economy that presumes that the greatest happiness results from antagonistic competitiveness.In an attempt to demystify how self-focused individuals may nonetheless thrive in“the space between”,we will examine,with the help of Kant and Foucault,the Enlightenment Project that promoted maximal reasonableness(or what we are referring to as“healthy moral selves”)and then,with the help of Steven Pinker and Alisdair MacIntyre,explore the factors that have led to its seeming relatively recent failure.We will argue,following Iris Murdoch and others,that the best hope for its revitalization,and with it,democracy,lies,on one hand,with the debunking of counterfeit moral selves who use a“moral stance”to deliver what Frankfurt refers to as“bullshit”,and,on the other,with the reinvigoration of our understanding that healthy moral selves require a steady diet of engaging in“objective practical reasoning”with those who think differently,thereby potentially starving our fat relentless egos that are so often the source of divisiveness and,in so doing,become happy by becoming worthy of being happy.In animating the value of this goal,the hope is that the means,i.e.,education for the reinvigoration of practical reason(the forgotten twin of theoretical reason)and genuine deliberative dialogue across difference will become sufficiently attractive that it will energize democratic practice and dialogue to such an extent that democracy,as a form of government,may yet flourish despite the atomizing forces of capitalism.展开更多
One of the key research focuses in quantum annealing is the design and optimization of annealing schedules to enhance computational efficiency,enabling large-scale applications.QuantumZero(QZero)pioneered the integrat...One of the key research focuses in quantum annealing is the design and optimization of annealing schedules to enhance computational efficiency,enabling large-scale applications.QuantumZero(QZero)pioneered the integration of Monte Carlo Tree Search(MCTS)with neural networks to autonomously design annealing schedules within a hybrid quantum-classical framework.This approach is distinguished by its ability to enhance the performance of Monte Carlo Tree Search through the integration of neural networks,enabling the efficient design of annealing paths even with limited annealing time.The paper presents an optimized QZero method based on intuitive reasoning theory and MindSpore,which further enhances QZero’s ability to conserve computational resources and resist noise.In terms of learning efficiency,the optimized QZero algorithm improves the convergence speed of the neural network by 93%compared to the original algorithm.Notably,the average number of quantum annealing queries required to achieve 99%fidelity is reduced by 45.09%.Regarding noise resistance,the optimized QZero algorithm requires 34.27%fewer quantum annealing queries to reach 99%fidelity compared to the original algorithm.The optimized QZero algorithm demonstrates strong competitiveness in optimizing quantum annealing schedules.展开更多
At inference time,deep neural networks are susceptible to backdoor attacks,which can produce attackercontrolled outputs when inputs contain carefully crafted triggers.Existing defense methods often focus on specific a...At inference time,deep neural networks are susceptible to backdoor attacks,which can produce attackercontrolled outputs when inputs contain carefully crafted triggers.Existing defense methods often focus on specific attack types or incur high costs,such as data cleaning or model fine-tuning.In contrast,we argue that it is possible to achieve effective and generalizable defense without removing triggers or incurring high model-cleaning costs.Fromthe attacker’s perspective and based on characteristics of vulnerable neuron activation anomalies,we propose an Adaptive Feature Injection(AFI)method for black-box backdoor detection.AFI employs a pre-trained image encoder to extract multi-level deep features and constructs a dynamic weight fusionmechanism for precise identification and interception of poisoned samples.Specifically,we select the control samples with the largest feature differences fromthe clean dataset via feature-space analysis,and generate blended sample pairs with the test sample using dynamic linear interpolation.The detection statistic is computed by measuring the divergence G(x)in model output responses.We systematically evaluate the effectiveness of AFI against representative backdoor attacks,including BadNets,Blend,WaNet,and IAB,on three benchmark datasets:MNIST,CIFAR-10,and ImageNet.Experimental results show that AFI can effectively detect poisoned samples,achieving average detection rates of 95.20%,94.15%,and 86.49%on these datasets,respectively.Compared with existing methods,AFI demonstrates strong cross-domain generalization ability and robustness to unknown attacks.展开更多
Planning in lexical-prior-free environments presents a fundamental challenge for evaluating whether large language models(LLMs)possess genuine structural reasoning capabilities beyond lexical memorization.When predica...Planning in lexical-prior-free environments presents a fundamental challenge for evaluating whether large language models(LLMs)possess genuine structural reasoning capabilities beyond lexical memorization.When predicates and action names are replaced with semantically irrelevant random symbols while preserving logical structures,existing direct generation approaches exhibit severe performance degradation.This paper proposes a symbol-agnostic closed-loop planning pipeline that enables models to construct executable plans through systematic validation and iterative refinement.The system implements a complete generate-verify-repair cycle through six core processing components:semantic comprehension extracts structural constraints,language planner generates text plans,symbol translator performs structure-preserving mapping,consistency checker conducts static screening,Stanford Research Institute Problem Solver(STRIPS)simulator executes step-by-step validation,and VAL(Validator)provides semantic verification.A repair controller orchestrates four targeted strategies addressing typical failure patterns including first-step precondition errors andmid-segment statemaintenance issues.Comprehensive evaluation on PlanBench Mystery Blocksworld demonstrates substantial improvements over baseline approaches across both language models and reasoning models.Ablation studies confirm that each architectural component contributes non-redundantly to overall effectiveness,with targeted repair providing the largest impact,followed by deep constraint extraction and stepwise validation,demonstrating that superior performance emerges from synergistic integration of these mechanisms rather than any single dominant factor.Analysis reveals distinct failure patterns betweenmodel types—languagemodels struggle with local precondition satisfaction while reasoning models face global goal achievement challenges—yet the validation-driven mechanism successfully addresses these diverse weaknesses.A particularly noteworthy finding is the convergence of final success rates across models with varying intrinsic capabilities,suggesting that systematic validation and repair mechanisms play a more decisive role than raw model capacity in lexical-prior-free scenarios.This work establishes a rigorous evaluation framework incorporating statistical significance testing and mechanistic failure analysis,providingmethodological contributions for fair assessment and practical insights into building reliable planning systems under extreme constraint conditions.展开更多
Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classificati...Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classification.This limitation becomes critical when hidden secondary cues—potentially more meaningful than the visualized ones—remain undiscovered.This study introduces CasCAM(Cascaded Class Activation Mapping)to address this fundamental limitation through counterfactual reasoning.By asking“if this dominant cue were absent,what other evidence would the model use?”,CasCAM progressively masks the most salient features and systematically uncovers the hierarchy of classification evidence hidden beneath them.Experimental results demonstrate that CasCAM effectively discovers the full spectrum of reasoning evidence and can be universally applied with nine existing interpretation methods.展开更多
In recent years bird hunting has become a topical issue in public debate in Poland.Some organizations and communities actively engaged in nature conservation efforts propose a complete ban on hunting,arguing it needs ...In recent years bird hunting has become a topical issue in public debate in Poland.Some organizations and communities actively engaged in nature conservation efforts propose a complete ban on hunting,arguing it needs to be introduced for ethical reasons and to ensure a more effective species protection.This discussion frequently involves cases when hunters break the law,illustrated by photographs published on social media,documenting culling of legally protected species.In response hunters mention the need to verify the sources and the context for published materials.展开更多
Artificial intelligence has experienced a significant boom with the emergence of agentic AI,where autonomous agents are increasingly replacing human intervention,enabling systems to perceive,reason,and act independent...Artificial intelligence has experienced a significant boom with the emergence of agentic AI,where autonomous agents are increasingly replacing human intervention,enabling systems to perceive,reason,and act independently to achieve specific goals.Despite its transformative potential,comprehensive information on agentic AI remains scarce in the literature.This paper provides the first comprehensive review of agentic AI,focusing on its evolution and three core aspects:patterns,types,and environments.The evolution of agentic AI is traced through five phases to the current era of multi-modal and collaborative agents,driven by advancements in reinforcement learning,neural networks,and large language models(LLMs).Five key patterns:tool use,reflection,ReAct,planning,and multi-agent collaboration(MAC)define how agentic AI systems interact and process tasks.These systems are categorized into seven categories,each tailored for specific operational styles and autonomy in decision making.The environments in which these agents operate are classified as static,dynamic,fully observable,partially observable,deterministic,stochastic,single-agent,and multiagent,emphasizing the impact of environmental complexity on agent behavior.Agentic AI has revolutionized systems through autonomous decision making and resource optimization,yet challenges persist in aligning AI with human values,ensuring adaptability,and addressing ethical constraints.Future research focuses on multidomain agents,human–AI collaboration,and self-improving systems.This work provides researchers,practitioners,and policymakers with a structured approach to understanding and advancing the rapidly evolving landscape of agentic AI systems.展开更多
Large LanguageModels(LLMs)are increasingly utilized for semantic understanding and reasoning,yet their use in sensitive settings is limited by privacy concerns.This paper presents In-Mig,a mobile-agent architecture th...Large LanguageModels(LLMs)are increasingly utilized for semantic understanding and reasoning,yet their use in sensitive settings is limited by privacy concerns.This paper presents In-Mig,a mobile-agent architecture that integrates LLM reasoning within agents that can migrate across organizational venues.Unlike centralized approaches,In-Mig performs reasoning in situ,ensuring that raw data remains within institutional boundaries while allowing for cross-venue synthesis.The architecture features a policy-scoped memory model,utility-driven route planning,and cryptographic trust enforcement.Aprototype using JADE for mobility and quantizedMistral-7B demonstrates practical feasibility.Evaluation across various scenarios shows that In-Mig achieves 92%similarity to centralized baselines,confirming its utility and strong privacy guarantees.These results suggest that migrating,privacy-preserving LLM agents can effectively support decentralized reasoning in trust-sensitive domains.展开更多
Knowledge graphs often suffer from sparsity and incompleteness.Knowledge graph reasoning is an effective way to address these issues.Unlike static knowledge graph reasoning,which is invariant over time,dynamic knowled...Knowledge graphs often suffer from sparsity and incompleteness.Knowledge graph reasoning is an effective way to address these issues.Unlike static knowledge graph reasoning,which is invariant over time,dynamic knowledge graph reasoning is more challenging due to its temporal nature.In essence,within each time step in a dynamic knowledge graph,there exists structural dependencies among entities and relations,whereas between adjacent time steps,there exists temporal continuity.Based on these structural and temporal characteristics,we propose a model named“DKGR-DR”to learn distributed representations of entities and relations by combining recurrent neural networks and graph neural networks to capture structural dependencies and temporal continuity in DKGs.In addition,we construct a static attribute graph to represent entities’inherent properties.DKGR-DR is capable of modeling both dynamic and static aspects of entities,enabling effective entity prediction and relation prediction.We conduct experiments on ICEWS05-15,ICEWS18,and ICEWS14 to demonstrate that DKGR-DR achieves competitive performance.展开更多
Large models,exemplified by ChatGPT,have reached the pinnacle of contemporary artificial intelligence(AI).However,they are plagued by three inherent drawbacks:excessive training data and computing power consumption,su...Large models,exemplified by ChatGPT,have reached the pinnacle of contemporary artificial intelligence(AI).However,they are plagued by three inherent drawbacks:excessive training data and computing power consumption,susceptibility to catastrophic forgetting,and a deficiency in logical reasoning capabilities within black-box models.To address these challenges,we draw insights from human memory mechanisms to introduce“machine memory,”which we define as a storage structure formed by encoding external information into a machine-representable and computable format.Centered on machine memory,we propose the brand-new machine memory intelligence(M^(2)I)framework,which encompasses representation,learning,and reasoning modules and loops.We explore the key issues and recent advances in the four core aspects of M^(2)I,including neural mechanisms,associative representation,continual learning,and collaborative reasoning within machine memory.M^(2)I aims to liberate machine intelligence from the confines of data-centric neural networks and fundamentally break through the limitations of existing large models,driving a qualitative leap from weak to strong AI.展开更多
This paper proposed a new systematic approach-functional evidential reasoning model(FERM) for exploring hazardous chemical operational accidents under uncertainty. First, FERM was introduced to identify various causal...This paper proposed a new systematic approach-functional evidential reasoning model(FERM) for exploring hazardous chemical operational accidents under uncertainty. First, FERM was introduced to identify various causal factors and their performance changes in hazardous chemical operational accidents, along with determining the functional failure link relationships. Subsequently, FERM was employed to elucidate both qualitative and quantitative operational accident information within a unified framework, which could be regarded as the input of information fusion to obtain the fuzzy belief distribution of each cause factor. Finally, the derived risk values of the causal factors were ranked while constructing multi-level accident causation chains to unveil the weak links in system functionality and the primary roots of operational accidents. Using the specific case of the “1·15” major explosion and fire accident at Liaoning Panjin Haoye Chemical Co., Ltd., seven causal factors and their corresponding performance changes were identified. Additionally, five accident causation chains were uncovered based on the fuzzy joint distribution of the functional assessment level(FAL) and reliability distribution(RD),revealing an overall increase in risk along the accident evolution path. The research findings demonstrated that FERM enabled the effective characterization, rational quantification and accurate analysis of the inherent uncertainties in hazardous chemical operational accident risks from a systemic perspective.展开更多
Automatically answer math word problems is a challenging task in artificial intelligence.Previous solvers constructed mathematical expressions in sequence or binary tree.However,these approaches may suffer from the fo...Automatically answer math word problems is a challenging task in artificial intelligence.Previous solvers constructed mathematical expressions in sequence or binary tree.However,these approaches may suffer from the following issues:Models relying on such structures exhibit fixed-order reasoning(e.g.,left-to-right),limiting flexibility and increasing error susceptibility;prior models rely on autoregressive reasoning in a single pass,accumulating minor errors(e.g.,incorrect math symbols)during generation,resulting in reduced accuracy.To address the above issues,we emulate the human“check and modify”process in reasoning and propose a unified M-tree self-correction solver(UTSCSolver)by iterative inference with self-correction mechanism.First,we use an iterative,non-autoregressive process for generating mathematical expressions,free from fixed generation orders to handle complex and diverse problems.Additionally,we design a self-correction mechanism based on alternating execution between a generator and a discriminator.This module iteratively detects and rectifies errors in generated expressions,leveraging previous iteration information for subsequent generation guidance.Experimental results show that our UTSC-Solver outperforms traditional models in accuracy on two popular datasets,while it improves the interpretability of mathematical reasoning.展开更多
真题回顾(2024·海南·中考真题)A hug(拥抱)is a form of human touch that happens when two or more people hold each other closely.People hug for many different reasons in their lives.For example,if a child is sad...真题回顾(2024·海南·中考真题)A hug(拥抱)is a form of human touch that happens when two or more people hold each other closely.People hug for many different reasons in their lives.For example,if a child is sad,a parent may hug him or her to give comfort.Grown-ups may hug to show each other love.Friends may hug to show friendship.Members of a team may hug after winning a game to show happiness and encourage other team members.展开更多
In this paper,a reasoning enhancement method based on RGCN(Relational Graph Convolutional Network)is proposed to improve the detection capability of UAV(Unmanned Aerial Vehicle)on fast-moving military targets in urban...In this paper,a reasoning enhancement method based on RGCN(Relational Graph Convolutional Network)is proposed to improve the detection capability of UAV(Unmanned Aerial Vehicle)on fast-moving military targets in urban battlefield environments.By combining military images with the publicly available VisDrone2019 dataset,a new dataset called VisMilitary was built and multiple YOLO(You Only Look Once)models were tested on it.Due to the low confidence problem caused by fuzzy targets,the performance of traditional YOLO models on real battlefield images decreases significantly.Therefore,we propose an improved RGCN inference model,which improves the performance of the model in complex environments by optimizing the data processing and graph network architecture.Experimental results show that the proposed method achieves an improvement of 0.4%to 1.7%on mAP@0.50,which proves the effectiveness of the model in military target detection.The research of this paper provides a new technical path for UAV target detection in urban battlefield,and provides important enlightenment for the application of deep learning in military field.展开更多
基金the National Key Research and Development Program of China (2016YFB1000603)the National Natural Science Foundation of China (NSFC)(Grant No.61672377)and the Key Technology Research and Development Program of Tianjin (16YFZCGX00210).
文摘Ontology classification,the problem of computing the subsumption hierarchies for classes (atomic concepts),is a core reasoning service provided by Web Ontology Language (OWL)reasoners.Although general-purpose OWL 2 reasoners employ sophisticated optimizations for classification,they are still not efficient owing to the high complexity of tableau algorithms for expressive ontologies. Profile-specific OWL 2 EL reasoners are efficient;however, they become incomplete even if the ontology contains only a small number of axioms that are outside the OWL 2 EL fragment.In this paper,we present a technique that combines an OWL 2 EL reasoner with an OWL 2 reasoner for ontology classification of expressive SROIQ.To optimize the workload,we propose a task decomposition strategy for identifying the minimal non-EL subontology that contains only necessary axioms to ensure completeness.During the ontology classification,the bulk of the workload is delegated to an efficient OWL 2 EL reasoner and only the minimal non- EL subontology is handled by a less efficient OWL 2 reasoner.The proposed approach is implemented in a prototype ComR and experimental results show that our approach offers a substantial speedup in ontology classification.For the wellknown ontology NCI,the classification time is reduced by 96.9%(resp.83.7%)compared against the standard reasoner Pellet (resp.the modular reasoner MORe).
文摘Multimodal dialogue systems often fail to maintain coherent reasoning over extended conversations and suffer from hallucination due to limited context modeling capabilities.Current approaches struggle with crossmodal alignment,temporal consistency,and robust handling of noisy or incomplete inputs across multiple modalities.We propose Multi Agent-Chain of Thought(CoT),a novel multi-agent chain-of-thought reasoning framework where specialized agents for text,vision,and speech modalities collaboratively construct shared reasoning traces through inter-agent message passing and consensus voting mechanisms.Our architecture incorporates self-reflection modules,conflict resolution protocols,and dynamic rationale alignment to enhance consistency,factual accuracy,and user engagement.The framework employs a hierarchical attention mechanism with cross-modal fusion and implements adaptive reasoning depth based on dialogue complexity.Comprehensive evaluations on Situated Interactive Multi-Modal Conversations(SIMMC)2.0,VisDial v1.0,and newly introduced challenging scenarios demonstrate statistically significant improvements in grounding accuracy(p<0.01),chain-of-thought interpretability,and robustness to adversarial inputs compared to state-of-the-art monolithic transformer baselines and existing multi-agent approaches.
文摘A fairly precise vision of a healthy physical self can serve as motivation for undertaking the means to that end.The same cannot be said with regard to“healthy moral selves”.By definition,democracy is about living with others and,as we argue,a healthy moral self is one that lives well with others.However,precisely what“living with others”entails is ambiguous,particularly in a capitalist economy that presumes that the greatest happiness results from antagonistic competitiveness.In an attempt to demystify how self-focused individuals may nonetheless thrive in“the space between”,we will examine,with the help of Kant and Foucault,the Enlightenment Project that promoted maximal reasonableness(or what we are referring to as“healthy moral selves”)and then,with the help of Steven Pinker and Alisdair MacIntyre,explore the factors that have led to its seeming relatively recent failure.We will argue,following Iris Murdoch and others,that the best hope for its revitalization,and with it,democracy,lies,on one hand,with the debunking of counterfeit moral selves who use a“moral stance”to deliver what Frankfurt refers to as“bullshit”,and,on the other,with the reinvigoration of our understanding that healthy moral selves require a steady diet of engaging in“objective practical reasoning”with those who think differently,thereby potentially starving our fat relentless egos that are so often the source of divisiveness and,in so doing,become happy by becoming worthy of being happy.In animating the value of this goal,the hope is that the means,i.e.,education for the reinvigoration of practical reason(the forgotten twin of theoretical reason)and genuine deliberative dialogue across difference will become sufficiently attractive that it will energize democratic practice and dialogue to such an extent that democracy,as a form of government,may yet flourish despite the atomizing forces of capitalism.
基金supported by the Defense Innovation Special Zone Project and CAAI-Huawei MindSpore Open Fund.
文摘One of the key research focuses in quantum annealing is the design and optimization of annealing schedules to enhance computational efficiency,enabling large-scale applications.QuantumZero(QZero)pioneered the integration of Monte Carlo Tree Search(MCTS)with neural networks to autonomously design annealing schedules within a hybrid quantum-classical framework.This approach is distinguished by its ability to enhance the performance of Monte Carlo Tree Search through the integration of neural networks,enabling the efficient design of annealing paths even with limited annealing time.The paper presents an optimized QZero method based on intuitive reasoning theory and MindSpore,which further enhances QZero’s ability to conserve computational resources and resist noise.In terms of learning efficiency,the optimized QZero algorithm improves the convergence speed of the neural network by 93%compared to the original algorithm.Notably,the average number of quantum annealing queries required to achieve 99%fidelity is reduced by 45.09%.Regarding noise resistance,the optimized QZero algorithm requires 34.27%fewer quantum annealing queries to reach 99%fidelity compared to the original algorithm.The optimized QZero algorithm demonstrates strong competitiveness in optimizing quantum annealing schedules.
基金supported by the National Natural Science Foundation of China Grant(No.61972133)Project of Leading Talents in Science and Technology Innovation for Thousands of People Plan in Henan Province Grant(No.204200510021)the Key Research and Development Plan Special Project of Henan Province Grant(No.241111211400).
文摘At inference time,deep neural networks are susceptible to backdoor attacks,which can produce attackercontrolled outputs when inputs contain carefully crafted triggers.Existing defense methods often focus on specific attack types or incur high costs,such as data cleaning or model fine-tuning.In contrast,we argue that it is possible to achieve effective and generalizable defense without removing triggers or incurring high model-cleaning costs.Fromthe attacker’s perspective and based on characteristics of vulnerable neuron activation anomalies,we propose an Adaptive Feature Injection(AFI)method for black-box backdoor detection.AFI employs a pre-trained image encoder to extract multi-level deep features and constructs a dynamic weight fusionmechanism for precise identification and interception of poisoned samples.Specifically,we select the control samples with the largest feature differences fromthe clean dataset via feature-space analysis,and generate blended sample pairs with the test sample using dynamic linear interpolation.The detection statistic is computed by measuring the divergence G(x)in model output responses.We systematically evaluate the effectiveness of AFI against representative backdoor attacks,including BadNets,Blend,WaNet,and IAB,on three benchmark datasets:MNIST,CIFAR-10,and ImageNet.Experimental results show that AFI can effectively detect poisoned samples,achieving average detection rates of 95.20%,94.15%,and 86.49%on these datasets,respectively.Compared with existing methods,AFI demonstrates strong cross-domain generalization ability and robustness to unknown attacks.
基金supported by the Information,Production and Systems Research Center,Waseda University,and partly supported by the Future Robotics Organization,Waseda Universitythe Humanoid Robotics Institute,Waseda University,under the Humanoid Project+1 种基金the Waseda University Grant for Special Research Projects(grant numbers 2024C-518 and 2025E-027)was partly executed under the cooperation of organization between Kioxia Corporation andWaseda University.
文摘Planning in lexical-prior-free environments presents a fundamental challenge for evaluating whether large language models(LLMs)possess genuine structural reasoning capabilities beyond lexical memorization.When predicates and action names are replaced with semantically irrelevant random symbols while preserving logical structures,existing direct generation approaches exhibit severe performance degradation.This paper proposes a symbol-agnostic closed-loop planning pipeline that enables models to construct executable plans through systematic validation and iterative refinement.The system implements a complete generate-verify-repair cycle through six core processing components:semantic comprehension extracts structural constraints,language planner generates text plans,symbol translator performs structure-preserving mapping,consistency checker conducts static screening,Stanford Research Institute Problem Solver(STRIPS)simulator executes step-by-step validation,and VAL(Validator)provides semantic verification.A repair controller orchestrates four targeted strategies addressing typical failure patterns including first-step precondition errors andmid-segment statemaintenance issues.Comprehensive evaluation on PlanBench Mystery Blocksworld demonstrates substantial improvements over baseline approaches across both language models and reasoning models.Ablation studies confirm that each architectural component contributes non-redundantly to overall effectiveness,with targeted repair providing the largest impact,followed by deep constraint extraction and stepwise validation,demonstrating that superior performance emerges from synergistic integration of these mechanisms rather than any single dominant factor.Analysis reveals distinct failure patterns betweenmodel types—languagemodels struggle with local precondition satisfaction while reasoning models face global goal achievement challenges—yet the validation-driven mechanism successfully addresses these diverse weaknesses.A particularly noteworthy finding is the convergence of final success rates across models with varying intrinsic capabilities,suggesting that systematic validation and repair mechanisms play a more decisive role than raw model capacity in lexical-prior-free scenarios.This work establishes a rigorous evaluation framework incorporating statistical significance testing and mechanistic failure analysis,providingmethodological contributions for fair assessment and practical insights into building reliable planning systems under extreme constraint conditions.
基金supported by the Basic Science Research Program through the National Research Foundation of Korea(NRF),funded by the Ministry of Education(RS-2023-00249743).
文摘Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classification.This limitation becomes critical when hidden secondary cues—potentially more meaningful than the visualized ones—remain undiscovered.This study introduces CasCAM(Cascaded Class Activation Mapping)to address this fundamental limitation through counterfactual reasoning.By asking“if this dominant cue were absent,what other evidence would the model use?”,CasCAM progressively masks the most salient features and systematically uncovers the hierarchy of classification evidence hidden beneath them.Experimental results demonstrate that CasCAM effectively discovers the full spectrum of reasoning evidence and can be universally applied with nine existing interpretation methods.
文摘In recent years bird hunting has become a topical issue in public debate in Poland.Some organizations and communities actively engaged in nature conservation efforts propose a complete ban on hunting,arguing it needs to be introduced for ethical reasons and to ensure a more effective species protection.This discussion frequently involves cases when hunters break the law,illustrated by photographs published on social media,documenting culling of legally protected species.In response hunters mention the need to verify the sources and the context for published materials.
文摘Artificial intelligence has experienced a significant boom with the emergence of agentic AI,where autonomous agents are increasingly replacing human intervention,enabling systems to perceive,reason,and act independently to achieve specific goals.Despite its transformative potential,comprehensive information on agentic AI remains scarce in the literature.This paper provides the first comprehensive review of agentic AI,focusing on its evolution and three core aspects:patterns,types,and environments.The evolution of agentic AI is traced through five phases to the current era of multi-modal and collaborative agents,driven by advancements in reinforcement learning,neural networks,and large language models(LLMs).Five key patterns:tool use,reflection,ReAct,planning,and multi-agent collaboration(MAC)define how agentic AI systems interact and process tasks.These systems are categorized into seven categories,each tailored for specific operational styles and autonomy in decision making.The environments in which these agents operate are classified as static,dynamic,fully observable,partially observable,deterministic,stochastic,single-agent,and multiagent,emphasizing the impact of environmental complexity on agent behavior.Agentic AI has revolutionized systems through autonomous decision making and resource optimization,yet challenges persist in aligning AI with human values,ensuring adaptability,and addressing ethical constraints.Future research focuses on multidomain agents,human–AI collaboration,and self-improving systems.This work provides researchers,practitioners,and policymakers with a structured approach to understanding and advancing the rapidly evolving landscape of agentic AI systems.
文摘Large LanguageModels(LLMs)are increasingly utilized for semantic understanding and reasoning,yet their use in sensitive settings is limited by privacy concerns.This paper presents In-Mig,a mobile-agent architecture that integrates LLM reasoning within agents that can migrate across organizational venues.Unlike centralized approaches,In-Mig performs reasoning in situ,ensuring that raw data remains within institutional boundaries while allowing for cross-venue synthesis.The architecture features a policy-scoped memory model,utility-driven route planning,and cryptographic trust enforcement.Aprototype using JADE for mobility and quantizedMistral-7B demonstrates practical feasibility.Evaluation across various scenarios shows that In-Mig achieves 92%similarity to centralized baselines,confirming its utility and strong privacy guarantees.These results suggest that migrating,privacy-preserving LLM agents can effectively support decentralized reasoning in trust-sensitive domains.
基金supported byNationalNatural Science Foundation of China(GrantNos.62071098,U24B20128)Sichuan Science and Technology Program(Grant No.2022YFG0319).
文摘Knowledge graphs often suffer from sparsity and incompleteness.Knowledge graph reasoning is an effective way to address these issues.Unlike static knowledge graph reasoning,which is invariant over time,dynamic knowledge graph reasoning is more challenging due to its temporal nature.In essence,within each time step in a dynamic knowledge graph,there exists structural dependencies among entities and relations,whereas between adjacent time steps,there exists temporal continuity.Based on these structural and temporal characteristics,we propose a model named“DKGR-DR”to learn distributed representations of entities and relations by combining recurrent neural networks and graph neural networks to capture structural dependencies and temporal continuity in DKGs.In addition,we construct a static attribute graph to represent entities’inherent properties.DKGR-DR is capable of modeling both dynamic and static aspects of entities,enabling effective entity prediction and relation prediction.We conduct experiments on ICEWS05-15,ICEWS18,and ICEWS14 to demonstrate that DKGR-DR achieves competitive performance.
基金supported by the National Natural Science Foun-dation of China(62137002,62250009,62202367,82025020,and 82230072).
文摘Large models,exemplified by ChatGPT,have reached the pinnacle of contemporary artificial intelligence(AI).However,they are plagued by three inherent drawbacks:excessive training data and computing power consumption,susceptibility to catastrophic forgetting,and a deficiency in logical reasoning capabilities within black-box models.To address these challenges,we draw insights from human memory mechanisms to introduce“machine memory,”which we define as a storage structure formed by encoding external information into a machine-representable and computable format.Centered on machine memory,we propose the brand-new machine memory intelligence(M^(2)I)framework,which encompasses representation,learning,and reasoning modules and loops.We explore the key issues and recent advances in the four core aspects of M^(2)I,including neural mechanisms,associative representation,continual learning,and collaborative reasoning within machine memory.M^(2)I aims to liberate machine intelligence from the confines of data-centric neural networks and fundamentally break through the limitations of existing large models,driving a qualitative leap from weak to strong AI.
基金supported by the National Key Research&Development Program of China(2021YFB3301100)the National Natural Science Foundation of China(52004014)the Fundamental Research Funds for the Central Universities(ZY2406).
文摘This paper proposed a new systematic approach-functional evidential reasoning model(FERM) for exploring hazardous chemical operational accidents under uncertainty. First, FERM was introduced to identify various causal factors and their performance changes in hazardous chemical operational accidents, along with determining the functional failure link relationships. Subsequently, FERM was employed to elucidate both qualitative and quantitative operational accident information within a unified framework, which could be regarded as the input of information fusion to obtain the fuzzy belief distribution of each cause factor. Finally, the derived risk values of the causal factors were ranked while constructing multi-level accident causation chains to unveil the weak links in system functionality and the primary roots of operational accidents. Using the specific case of the “1·15” major explosion and fire accident at Liaoning Panjin Haoye Chemical Co., Ltd., seven causal factors and their corresponding performance changes were identified. Additionally, five accident causation chains were uncovered based on the fuzzy joint distribution of the functional assessment level(FAL) and reliability distribution(RD),revealing an overall increase in risk along the accident evolution path. The research findings demonstrated that FERM enabled the effective characterization, rational quantification and accurate analysis of the inherent uncertainties in hazardous chemical operational accident risks from a systemic perspective.
基金supported by the National Natural Science Foundation of China(62106244)the Fundamental Research Funds for the Central Universities(WK2150110021)the University Synergy Innovation Program of Anhui Province(GXXT-2022-042).
文摘Automatically answer math word problems is a challenging task in artificial intelligence.Previous solvers constructed mathematical expressions in sequence or binary tree.However,these approaches may suffer from the following issues:Models relying on such structures exhibit fixed-order reasoning(e.g.,left-to-right),limiting flexibility and increasing error susceptibility;prior models rely on autoregressive reasoning in a single pass,accumulating minor errors(e.g.,incorrect math symbols)during generation,resulting in reduced accuracy.To address the above issues,we emulate the human“check and modify”process in reasoning and propose a unified M-tree self-correction solver(UTSCSolver)by iterative inference with self-correction mechanism.First,we use an iterative,non-autoregressive process for generating mathematical expressions,free from fixed generation orders to handle complex and diverse problems.Additionally,we design a self-correction mechanism based on alternating execution between a generator and a discriminator.This module iteratively detects and rectifies errors in generated expressions,leveraging previous iteration information for subsequent generation guidance.Experimental results show that our UTSC-Solver outperforms traditional models in accuracy on two popular datasets,while it improves the interpretability of mathematical reasoning.
文摘真题回顾(2024·海南·中考真题)A hug(拥抱)is a form of human touch that happens when two or more people hold each other closely.People hug for many different reasons in their lives.For example,if a child is sad,a parent may hug him or her to give comfort.Grown-ups may hug to show each other love.Friends may hug to show friendship.Members of a team may hug after winning a game to show happiness and encourage other team members.
基金supported by the National Natural Science Foundation of China(61806024,62206257)the Jilin Province Science and Technology Development Plan Key Research and Development Project(20210204050YY)+1 种基金the Wuxi University Research Start-up Fund for Introduced Talents(2023r004,2023r006)Jiangsu Engineering Research Center of Hyperconvergence Application and Security of IoT Devices,Jiangsu Foreign Expert Workshop,Wuxi City Internet of Vehicles Key Laboratory.
文摘In this paper,a reasoning enhancement method based on RGCN(Relational Graph Convolutional Network)is proposed to improve the detection capability of UAV(Unmanned Aerial Vehicle)on fast-moving military targets in urban battlefield environments.By combining military images with the publicly available VisDrone2019 dataset,a new dataset called VisMilitary was built and multiple YOLO(You Only Look Once)models were tested on it.Due to the low confidence problem caused by fuzzy targets,the performance of traditional YOLO models on real battlefield images decreases significantly.Therefore,we propose an improved RGCN inference model,which improves the performance of the model in complex environments by optimizing the data processing and graph network architecture.Experimental results show that the proposed method achieves an improvement of 0.4%to 1.7%on mAP@0.50,which proves the effectiveness of the model in military target detection.The research of this paper provides a new technical path for UAV target detection in urban battlefield,and provides important enlightenment for the application of deep learning in military field.
