Selected work

Representative Papers

Spatial-Frequency-Aware Implicit Neural Representation of Multidimensional Signals via MLP-KAN Fusion

Oct 08, 2026

This study addresses the spectral bias in MLP-based implicit neural representations, which leads to inadequate reconstruction of high-frequency details. To overcome this limitation, we propose a spatial-frequency-aware framework that integrates MLPs with Kolmogorov–Arnold Networks (KANs) to achieve complementary frequency modeling. By incorporating the discrete wavelet transform, the input signal is decomposed into distinct frequency bands, and a band-separation regularization term is introduced to guide each branch toward specialized yet synergistic reconstruction. Extensive experiments demonstrate that the proposed method significantly improves reconstruction fidelity across multidimensional signal tasks, validating its cross-modal applicability and strong generalization capability for efficient and accurate signal representation.

0 citationsRead paper

Acmite: Mitigating Gender Bias in LLMs through Concept-Guided Mutual Information

Oct 01, 2026

This study addresses the limitations of existing LLM debiasing methods, which rely on explicit examples or fixed substitutions and struggle to capture cross-contextual biases or model the statistical dependence between outputs and biases. We propose Acmite, a lightweight concept-guided framework that structures stereotypes into semantic concepts, filters them via Maximum Marginal Relevance, and achieves targeted debiasing through mutual information minimization. Furthermore, it introduces a conditional triggering mechanism that activates LoRA adapters only upon bias detection, effectively balancing debiasing performance with the preservation of general capabilities. Experimental results demonstrate that Acmite significantly mitigates gender bias on benchmarks such as BBQ while maintaining competitive performance on downstream tasks including ARC and GSM8K.

0 citationsRead paper

V2X-WAM: A Cooperative World Action Model for End-to-End Autonomous Driving

Sep 29, 2026

Existing cooperative driving methods lack explicit modeling of the future consequences of behaviors, limiting their planning capabilities in dynamic environments. This work proposes V2X-WAM, a model that tightly couples scene understanding, action generation, and future world reasoning to optimize end-to-end planning via closed-loop feedback. Key innovations include the first approach to compressing infrastructure information into quantized messages for substantially reduced communication overhead, the construction of reliability-aware spatiotemporal representations alongside a multimodal planner, and the establishment of a closed-loop interaction mechanism between actions and future occupancy and flow prediction. Evaluated on large-scale real-world datasets, the proposed model significantly improves planning accuracy, safety, and predictive performance.

0 citationsRead paper

Evolution of fairness in multi-objective reinforcement learning framework

Sep 28, 2026

This study addresses the limitations of traditional models in explaining the evolutionary mechanisms of fairness perception, which often overlook the multidimensionality of human decision-making. To this end, we propose a multi-objective reinforcement learning framework that introduces a fairness pressure coefficient and employs a dual-objective Q-learning algorithm to simulate the Ultimatum Game, capturing the dynamic trade-off between material payoffs and moral fairness. Our simulations reveal a reversal phenomenon of "tolerant" strategies under moderate pressure alongside preference competition mechanisms, validating the critical influence of fairness pressure on decision outcomes. By extending the reinforcement learning paradigm, this work offers a novel analytical perspective for understanding broader social behaviors.

0 citationsRead paper

Distance-KV: Exploiting Relative Distance for Efficient Long-Context Inference

Sep 26, 2026

This study addresses the memory surge and high latency challenges in long-context reasoning for large language models, noting that existing KV cache compression methods overlook the critical impact of relative distance on retrieval capability. We propose Distance-KV, which reveals for the first time that relative distance constitutes a core structural dimension for long-context retrieval. This method learns a static KV retention pattern by jointly optimizing across layers, attention heads, and relative distances offline on a frozen language model, enabling direct cache pruning during inference without online importance scoring. Experimental results demonstrate that Distance-KV outperforms the strongest baseline by 9.3 points on the RULER benchmark while reducing the memory footprint of Llama-3.1-8B by 65.4% and accelerating decoding speed by 1.66×.

0 citationsRead paper

Recent publications

Latest Papers

Spatial-Frequency-Aware Implicit Neural Representation of Multidimensional Signals via MLP-KAN Fusion

Oct 08, 2026

This study addresses the spectral bias in MLP-based implicit neural representations, which leads to inadequate reconstruction of high-frequency details. To overcome this limitation, we propose a spatial-frequency-aware framework that integrates MLPs with Kolmogorov–Arnold Networks (KANs) to achieve complementary frequency modeling. By incorporating the discrete wavelet transform, the input signal is decomposed into distinct frequency bands, and a band-separation regularization term is introduced to guide each branch toward specialized yet synergistic reconstruction. Extensive experiments demonstrate that the proposed method significantly improves reconstruction fidelity across multidimensional signal tasks, validating its cross-modal applicability and strong generalization capability for efficient and accurate signal representation.

0 citationsRead paper

Acmite: Mitigating Gender Bias in LLMs through Concept-Guided Mutual Information

Oct 01, 2026

This study addresses the limitations of existing LLM debiasing methods, which rely on explicit examples or fixed substitutions and struggle to capture cross-contextual biases or model the statistical dependence between outputs and biases. We propose Acmite, a lightweight concept-guided framework that structures stereotypes into semantic concepts, filters them via Maximum Marginal Relevance, and achieves targeted debiasing through mutual information minimization. Furthermore, it introduces a conditional triggering mechanism that activates LoRA adapters only upon bias detection, effectively balancing debiasing performance with the preservation of general capabilities. Experimental results demonstrate that Acmite significantly mitigates gender bias on benchmarks such as BBQ while maintaining competitive performance on downstream tasks including ARC and GSM8K.

0 citationsRead paper

V2X-WAM: A Cooperative World Action Model for End-to-End Autonomous Driving

Sep 29, 2026

Existing cooperative driving methods lack explicit modeling of the future consequences of behaviors, limiting their planning capabilities in dynamic environments. This work proposes V2X-WAM, a model that tightly couples scene understanding, action generation, and future world reasoning to optimize end-to-end planning via closed-loop feedback. Key innovations include the first approach to compressing infrastructure information into quantized messages for substantially reduced communication overhead, the construction of reliability-aware spatiotemporal representations alongside a multimodal planner, and the establishment of a closed-loop interaction mechanism between actions and future occupancy and flow prediction. Evaluated on large-scale real-world datasets, the proposed model significantly improves planning accuracy, safety, and predictive performance.

0 citationsRead paper

Evolution of fairness in multi-objective reinforcement learning framework

Sep 28, 2026

This study addresses the limitations of traditional models in explaining the evolutionary mechanisms of fairness perception, which often overlook the multidimensionality of human decision-making. To this end, we propose a multi-objective reinforcement learning framework that introduces a fairness pressure coefficient and employs a dual-objective Q-learning algorithm to simulate the Ultimatum Game, capturing the dynamic trade-off between material payoffs and moral fairness. Our simulations reveal a reversal phenomenon of "tolerant" strategies under moderate pressure alongside preference competition mechanisms, validating the critical influence of fairness pressure on decision outcomes. By extending the reinforcement learning paradigm, this work offers a novel analytical perspective for understanding broader social behaviors.

0 citationsRead paper

Distance-KV: Exploiting Relative Distance for Efficient Long-Context Inference

Sep 26, 2026

This study addresses the memory surge and high latency challenges in long-context reasoning for large language models, noting that existing KV cache compression methods overlook the critical impact of relative distance on retrieval capability. We propose Distance-KV, which reveals for the first time that relative distance constitutes a core structural dimension for long-context retrieval. This method learns a static KV retention pattern by jointly optimizing across layers, attention heads, and relative distances offline on a frozen language model, enabling direct cache pruning during inference without online importance scoring. Experimental results demonstrate that Distance-KV outperforms the strongest baseline by 9.3 points on the RULER benchmark while reducing the memory footprint of Llama-3.1-8B by 65.4% and accelerating decoding speed by 1.66×.

0 citationsRead paper