Selected work

Representative Papers

Query Expansion and Key Specialization in Transformer Attention Geometry

Sep 28, 2026

This study investigates the divergent geometric evolution of query (Q) and key (K) projections during Transformer training and its implications for the attention mechanism. By systematically monitoring the training trajectories of small GPT models through participation ratio analysis, effective dimensionality tracking, and spectral control experiments, this work reveals, for the first time, the dynamic evolution of Q/K geometric asymmetry: the effective dimensionality of Q expands while that of K contracts. Furthermore, it establishes a causal link between this asymmetry and changes in attention entropy. Experimental results confirm that spectral contraction in K directly drives the sharpening of attention distributions. The findings also characterize the universality of these geometric trends during early training and their subsequent decay in later stages, offering a novel perspective for understanding attention mechanisms.

0 citationsRead paper

FRAC-MAS: A Safe and Explainable Multi-Agent System for Fracture Diagnosis

Aug 22, 2026

本文提出FRAC-MAS系统，通过结合视觉模型和多代理工作流解决骨骨折诊断的可解释性和安全性问题，提高了诊断的准确性与临床适用性。

0 citationsRead paper

Advanced Encryption Technique for Multimedia Data Using Sudoku-Based Algorithms for Enhanced Security

Jan 15, 2026

This work proposes a unified block cipher framework based on Sudoku puzzles to address the vulnerability of multimedia data during transmission and the limitations of conventional encryption methods in achieving both robust security and multimodal compatibility. For the first time, a timestamp-driven dynamic key mechanism is integrated into the Sudoku-based encryption scheme, enabling efficient and secure encryption of images, audio, and video within a single architecture. The approach combines Sudoku permutation with XOR substitution to form a lightweight block cipher. Experimental results demonstrate that the proposed method achieves near-perfect NPCR (close to 100%) for image encryption and an SNR exceeding 60 dB for audio encryption, significantly enhancing resistance against brute-force and differential attacks.

0 citationsRead paper

Color, Sentiment, and Structure: A Comparative Study of Instagram Marketing Across Economies

Dec 20, 2025

This study investigates how aesthetic features (dominant hue, textual sentiment) and structural macroeconomic variables (GDP, population, obesity prevalence) jointly shape consumer engagement with global food brands on Instagram, and how these effects vary across economies. Employing multimodal analysis—HSV color space extraction and VADER sentiment scoring—alongside hierarchical regression modeling, it is the first to systematically integrate visual, linguistic, and macro-level predictors. Results reveal distinct cross-national patterns: in developing economies, beige–green palettes correlate with higher engagement, and GDP positively predicts interaction; in developed economies, population size enhances engagement, whereas GDP negatively moderates attention. Obesity prevalence exhibits significant regional heterogeneity in its effect on likes and comments. The findings uncover nonlinear and directionally opposing moderation effects, advancing theoretical understanding of digital marketing localization and offering empirically grounded guidance for region-specific campaign design.

0 citationsRead paper

Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning

Sep 08, 2025

This study addresses the low efficiency and unstable performance of open-source large language models (LLMs) on mathematical reasoning tasks. We propose the first cross-architecture, systematic parameter optimization framework, jointly tuning temperature (0.1–0.5), reasoning steps (4–12), planning cycles (1–4), and nucleus sampling threshold (0.85–0.98), while incorporating stochasticity control and dynamic depth adjustment. The framework achieves 100% optimization success across five state-of-the-art models: Qwen2.5-72B, Llama-3.1-70B, DeepSeek-V3, Mixtral-8x22B, and Yi-Lightning. Experiments demonstrate an average 29.4% reduction in computational cost, a 23.9% increase in inference speed, a 98% accuracy for DeepSeek-V3, and peak token efficiency for Mixtral-8x22B (361.5 tokens per correct response). The work establishes a reusable, standardized optimization paradigm and plug-and-play configuration protocol, providing both theoretical foundations and practical engineering guidance for efficient LLM deployment in mathematical reasoning.

0 citationsRead paper

Recent publications

Latest Papers

Query Expansion and Key Specialization in Transformer Attention Geometry

Sep 28, 2026

This study investigates the divergent geometric evolution of query (Q) and key (K) projections during Transformer training and its implications for the attention mechanism. By systematically monitoring the training trajectories of small GPT models through participation ratio analysis, effective dimensionality tracking, and spectral control experiments, this work reveals, for the first time, the dynamic evolution of Q/K geometric asymmetry: the effective dimensionality of Q expands while that of K contracts. Furthermore, it establishes a causal link between this asymmetry and changes in attention entropy. Experimental results confirm that spectral contraction in K directly drives the sharpening of attention distributions. The findings also characterize the universality of these geometric trends during early training and their subsequent decay in later stages, offering a novel perspective for understanding attention mechanisms.

0 citationsRead paper

FRAC-MAS: A Safe and Explainable Multi-Agent System for Fracture Diagnosis

Aug 22, 2026

本文提出FRAC-MAS系统，通过结合视觉模型和多代理工作流解决骨骨折诊断的可解释性和安全性问题，提高了诊断的准确性与临床适用性。

0 citationsRead paper

Advanced Encryption Technique for Multimedia Data Using Sudoku-Based Algorithms for Enhanced Security

Jan 15, 2026

This work proposes a unified block cipher framework based on Sudoku puzzles to address the vulnerability of multimedia data during transmission and the limitations of conventional encryption methods in achieving both robust security and multimodal compatibility. For the first time, a timestamp-driven dynamic key mechanism is integrated into the Sudoku-based encryption scheme, enabling efficient and secure encryption of images, audio, and video within a single architecture. The approach combines Sudoku permutation with XOR substitution to form a lightweight block cipher. Experimental results demonstrate that the proposed method achieves near-perfect NPCR (close to 100%) for image encryption and an SNR exceeding 60 dB for audio encryption, significantly enhancing resistance against brute-force and differential attacks.

0 citationsRead paper

Color, Sentiment, and Structure: A Comparative Study of Instagram Marketing Across Economies

Dec 20, 2025

This study investigates how aesthetic features (dominant hue, textual sentiment) and structural macroeconomic variables (GDP, population, obesity prevalence) jointly shape consumer engagement with global food brands on Instagram, and how these effects vary across economies. Employing multimodal analysis—HSV color space extraction and VADER sentiment scoring—alongside hierarchical regression modeling, it is the first to systematically integrate visual, linguistic, and macro-level predictors. Results reveal distinct cross-national patterns: in developing economies, beige–green palettes correlate with higher engagement, and GDP positively predicts interaction; in developed economies, population size enhances engagement, whereas GDP negatively moderates attention. Obesity prevalence exhibits significant regional heterogeneity in its effect on likes and comments. The findings uncover nonlinear and directionally opposing moderation effects, advancing theoretical understanding of digital marketing localization and offering empirically grounded guidance for region-specific campaign design.

0 citationsRead paper

Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning

Sep 08, 2025

This study addresses the low efficiency and unstable performance of open-source large language models (LLMs) on mathematical reasoning tasks. We propose the first cross-architecture, systematic parameter optimization framework, jointly tuning temperature (0.1–0.5), reasoning steps (4–12), planning cycles (1–4), and nucleus sampling threshold (0.85–0.98), while incorporating stochasticity control and dynamic depth adjustment. The framework achieves 100% optimization success across five state-of-the-art models: Qwen2.5-72B, Llama-3.1-70B, DeepSeek-V3, Mixtral-8x22B, and Yi-Lightning. Experiments demonstrate an average 29.4% reduction in computational cost, a 23.9% increase in inference speed, a 98% accuracy for DeepSeek-V3, and peak token efficiency for Mixtral-8x22B (361.5 tokens per correct response). The work establishes a reusable, standardized optimization paradigm and plug-and-play configuration protocol, providing both theoretical foundations and practical engineering guidance for efficient LLM deployment in mathematical reasoning.

0 citationsRead paper