Selected work

Representative Papers

Classical codes violate the conjectured square-root bound for quantum random access codes

Jul 17, 2026

This work investigates whether quantum random access codes satisfy the conjectured square-root bound $p \leq (1 + \sqrt{m/n})/2$. By embedding classical random access codes with private randomness into quantum schemes using diagonal encoding states and commuting POVM measurements, the authors construct the first classical counterexamples that violate this bound, revealing that classical coding rates are key to the separation from the quantum limit. Leveraging spectral properties of decoding measurements, they establish a new restricted bound and prove that for any fixed $p \in (1/2, 1)$, counterexamples exist when the input length is sufficiently large. Moreover, they achieve optimal logarithmic qubit scaling when the recovery bias is $\sqrt{\log_2 n / n}$.

0 citationsRead paper

CLLMRec: LLM-powered Cognitive-Aware Concept Recommendation via Semantic Alignment and Prerequisite Knowledge Distillation

Nov 21, 2025

Personalized concept recommendation in MOOCs is hindered by reliance on manually constructed knowledge graphs, which are labor-intensive and scarce in high quality. Method: This paper proposes a cognitive-aware recommendation framework that eliminates the need for structured knowledge graphs. It leverages large language models (LLMs) to automatically extract prerequisite relationships among concepts using their inherent world knowledge; achieves unified concept representation via semantic alignment; transfers LLM-derived priors to a lightweight student model through teacher–student knowledge distillation; and integrates deep knowledge tracing to model learners’ real-time cognitive states, enhanced by a fine-grained ranking mechanism for personalization. Results: Experiments on two real-world MOOC datasets demonstrate significant improvements over state-of-the-art baselines in recommendation accuracy, cognitive adaptability, and robustness.

0 citationsRead paper

Guided Depth Map Super-Resolution via Multi-Scale Fusion U-shaped Mamba Network

Jul 31, 2025

To address the trade-off between weak global modeling capability and high computational cost in depth map super-resolution, this paper proposes a color-image-guided multi-scale Mamba-U network. Methodologically, it introduces the Mamba state-space model—previously unexplored in depth SR—into a cross-modal multi-scale U-shaped architecture, integrating residual dense channel attention with a convolution-state-space hybrid modeling mechanism to efficiently capture long-range dependencies while preserving fine local details. Key contributions include: (1) a lightweight cross-modal sequence modeling strategy that significantly reduces computational complexity; and (2) unified optimization for global contextual awareness and local feature enhancement. Evaluated on NYUv2 and KITTI, our method achieves state-of-the-art reconstruction accuracy with fewer parameters—improving PSNR by 0.82 dB and inference speed by 37% on 4× super-resolution tasks.

0 citationsRead paper

Recent publications

Latest Papers

Classical codes violate the conjectured square-root bound for quantum random access codes

Jul 17, 2026

This work investigates whether quantum random access codes satisfy the conjectured square-root bound $p \leq (1 + \sqrt{m/n})/2$. By embedding classical random access codes with private randomness into quantum schemes using diagonal encoding states and commuting POVM measurements, the authors construct the first classical counterexamples that violate this bound, revealing that classical coding rates are key to the separation from the quantum limit. Leveraging spectral properties of decoding measurements, they establish a new restricted bound and prove that for any fixed $p \in (1/2, 1)$, counterexamples exist when the input length is sufficiently large. Moreover, they achieve optimal logarithmic qubit scaling when the recovery bias is $\sqrt{\log_2 n / n}$.

0 citationsRead paper

CLLMRec: LLM-powered Cognitive-Aware Concept Recommendation via Semantic Alignment and Prerequisite Knowledge Distillation

Nov 21, 2025

Personalized concept recommendation in MOOCs is hindered by reliance on manually constructed knowledge graphs, which are labor-intensive and scarce in high quality. Method: This paper proposes a cognitive-aware recommendation framework that eliminates the need for structured knowledge graphs. It leverages large language models (LLMs) to automatically extract prerequisite relationships among concepts using their inherent world knowledge; achieves unified concept representation via semantic alignment; transfers LLM-derived priors to a lightweight student model through teacher–student knowledge distillation; and integrates deep knowledge tracing to model learners’ real-time cognitive states, enhanced by a fine-grained ranking mechanism for personalization. Results: Experiments on two real-world MOOC datasets demonstrate significant improvements over state-of-the-art baselines in recommendation accuracy, cognitive adaptability, and robustness.

0 citationsRead paper

Guided Depth Map Super-Resolution via Multi-Scale Fusion U-shaped Mamba Network

Jul 31, 2025

To address the trade-off between weak global modeling capability and high computational cost in depth map super-resolution, this paper proposes a color-image-guided multi-scale Mamba-U network. Methodologically, it introduces the Mamba state-space model—previously unexplored in depth SR—into a cross-modal multi-scale U-shaped architecture, integrating residual dense channel attention with a convolution-state-space hybrid modeling mechanism to efficiently capture long-range dependencies while preserving fine local details. Key contributions include: (1) a lightweight cross-modal sequence modeling strategy that significantly reduces computational complexity; and (2) unified optimization for global contextual awareness and local feature enhancement. Evaluated on NYUv2 and KITTI, our method achieves state-of-the-art reconstruction accuracy with fewer parameters—improving PSNR by 0.82 dB and inference speed by 37% on 4× super-resolution tasks.

0 citationsRead paper