Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme
本文针对越南2025年高考非线性评分系统下部分正确答案得分问题，提出THPT-Ladder基准来准确评估语言模型表现。
0 citationsRead paper
Selected work
Representative Papers
Provable Parameter-Free Fixed-Point Algorithms with Linear Convergence Rates
This work addresses the challenge of computing fixed points of contractive mappings without requiring prior knowledge of the contraction factor or manual hyperparameter tuning. We propose a fully adaptive Halpern-type algorithm that operates without line searches, bisection procedures, or any user-specified parameters, automatically exploiting the inherent contractiveness of the mapping to achieve explicit linear convergence even in the absence of a priori estimates of the contraction constant. Theoretical analysis establishes linear convergence rates both in terms of fixed-point residuals and distance to the solution, with an iteration complexity of 𝒪(ε⁻¹ln(ε⁻¹)) for solving cocoercive equations. By integrating Tikhonov regularization and Nesterov acceleration, we further extend the algorithm’s applicability. Numerical experiments confirm its superiority over existing adaptive methods, offering strong theoretical guarantees alongside low computational overhead.
0 citationsRead paper
VIVID: A Culturally Grounded Benchmark Exposing the Figurative Language Gap in Vietnamese NLP
This study addresses the significant limitations of large language models in understanding Vietnamese figurative expressions—such as idioms and proverbs—that are deeply rooted in cultural context. To this end, the authors introduce VIVID, the first systematic evaluation benchmark for Vietnamese figurative language, comprising 1,636 annotated expressions labeled for complexity and semantic themes. They propose an evaluation framework integrating generative and discriminative tasks, augmented with a human-validated, aspect-level LLM-as-a-Judge mechanism (Cohen’s κ = 0.792). Experiments on eight state-of-the-art models reveal that Vietnamese-specific models substantially underperform multilingual counterparts (e.g., VinaLLaMA-7B scores 0.13 versus GPT-4o’s 2.46 on a 5-point scale), with all models scoring below 50% of the maximum. Notably, few-shot prompting degrades GPT-4o’s performance due to stylistic overfitting, underscoring a systemic deficiency in cultural pragmatic comprehension.
0 citationsRead paper
SeCo-SBIR: Semantically Consistent Prompt Learning for Zero-Shot Sketch-Based Image Retrieval
This work addresses the challenge of balancing cross-domain adaptation and zero-shot generalization in zero-shot sketch-based image retrieval (ZS-SBIR). To this end, the authors propose a semantic-consistent prompt learning framework that effectively adapts the CLIP model through a text-guided intermediate representation injection mechanism and a perturbation-based asymmetric consistency constraint. This approach preserves CLIP’s inherent generalization capability while significantly enhancing retrieval performance. The framework incorporates learnable prompt vectors, a cross-modal coupling function, and a lightweight adapter module, optimized jointly via a multi-objective loss combining triplet, NT-Xent, and classification losses. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance across all three standard ZS-SBIR benchmarks—within-category, generalized, and cross-dataset settings.
0 citationsRead paper
SpatialQuery: Benchmarking Geometry-Grounded Multi-Instance Spatial Reasoning in Vision-Language Models
This work addresses the poor performance of existing vision-language models in metric spatial reasoning tasks involving multiple instances of the same category, where accurately identifying the nearest neighbor and its ground-truth distance remains challenging. The authors propose a training-free framework that enhances spatial reasoning without modifying model architecture or fine-tuning. It first applies Scene Cubifying to transform objects in an image into canonical bird’s-eye-view blocks of uniform size, each encoded with category information. This representation is then coupled with an Uncertainty-Aware Chain-of-Thought prompting mechanism to improve reasoning fidelity. Evaluated on the SPATIALQUERY-1M benchmark, the method achieves state-of-the-art results with Qwen3-VL-8B, yielding a Floor-MAE of 0.259 meters, an Uncertainty-Aware Accuracy@0.3m of 90.5%, and a nearest-neighbor identification accuracy of 84.18%, outperforming both specialized models and leading closed-source systems.
0 citationsRead paper
Recent publications
Latest Papers
Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme
本文针对越南2025年高考非线性评分系统下部分正确答案得分问题，提出THPT-Ladder基准来准确评估语言模型表现。
0 citationsRead paper
Provable Parameter-Free Fixed-Point Algorithms with Linear Convergence Rates
This work addresses the challenge of computing fixed points of contractive mappings without requiring prior knowledge of the contraction factor or manual hyperparameter tuning. We propose a fully adaptive Halpern-type algorithm that operates without line searches, bisection procedures, or any user-specified parameters, automatically exploiting the inherent contractiveness of the mapping to achieve explicit linear convergence even in the absence of a priori estimates of the contraction constant. Theoretical analysis establishes linear convergence rates both in terms of fixed-point residuals and distance to the solution, with an iteration complexity of 𝒪(ε⁻¹ln(ε⁻¹)) for solving cocoercive equations. By integrating Tikhonov regularization and Nesterov acceleration, we further extend the algorithm’s applicability. Numerical experiments confirm its superiority over existing adaptive methods, offering strong theoretical guarantees alongside low computational overhead.
0 citationsRead paper
VIVID: A Culturally Grounded Benchmark Exposing the Figurative Language Gap in Vietnamese NLP
This study addresses the significant limitations of large language models in understanding Vietnamese figurative expressions—such as idioms and proverbs—that are deeply rooted in cultural context. To this end, the authors introduce VIVID, the first systematic evaluation benchmark for Vietnamese figurative language, comprising 1,636 annotated expressions labeled for complexity and semantic themes. They propose an evaluation framework integrating generative and discriminative tasks, augmented with a human-validated, aspect-level LLM-as-a-Judge mechanism (Cohen’s κ = 0.792). Experiments on eight state-of-the-art models reveal that Vietnamese-specific models substantially underperform multilingual counterparts (e.g., VinaLLaMA-7B scores 0.13 versus GPT-4o’s 2.46 on a 5-point scale), with all models scoring below 50% of the maximum. Notably, few-shot prompting degrades GPT-4o’s performance due to stylistic overfitting, underscoring a systemic deficiency in cultural pragmatic comprehension.
0 citationsRead paper
SeCo-SBIR: Semantically Consistent Prompt Learning for Zero-Shot Sketch-Based Image Retrieval
This work addresses the challenge of balancing cross-domain adaptation and zero-shot generalization in zero-shot sketch-based image retrieval (ZS-SBIR). To this end, the authors propose a semantic-consistent prompt learning framework that effectively adapts the CLIP model through a text-guided intermediate representation injection mechanism and a perturbation-based asymmetric consistency constraint. This approach preserves CLIP’s inherent generalization capability while significantly enhancing retrieval performance. The framework incorporates learnable prompt vectors, a cross-modal coupling function, and a lightweight adapter module, optimized jointly via a multi-objective loss combining triplet, NT-Xent, and classification losses. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance across all three standard ZS-SBIR benchmarks—within-category, generalized, and cross-dataset settings.
0 citationsRead paper
SpatialQuery: Benchmarking Geometry-Grounded Multi-Instance Spatial Reasoning in Vision-Language Models
This work addresses the poor performance of existing vision-language models in metric spatial reasoning tasks involving multiple instances of the same category, where accurately identifying the nearest neighbor and its ground-truth distance remains challenging. The authors propose a training-free framework that enhances spatial reasoning without modifying model architecture or fine-tuning. It first applies Scene Cubifying to transform objects in an image into canonical bird’s-eye-view blocks of uniform size, each encoded with category information. This representation is then coupled with an Uncertainty-Aware Chain-of-Thought prompting mechanism to improve reasoning fidelity. Evaluated on the SPATIALQUERY-1M benchmark, the method achieves state-of-the-art results with Qwen3-VL-8B, yielding a Floor-MAE of 0.259 meters, an Uncertainty-Aware Accuracy@0.3m of 90.5%, and a nearest-neighbor identification accuracy of 84.18%, outperforming both specialized models and leading closed-source systems.
0 citationsRead paper