Beam Search, Self-Consistency, and the Limits of Inference-Time Scaling for Grammar-Constrained Text-to-SQL in Small Language Models
研究探讨了在语法受限条件下，通过调整推理时计算量（如改变束搜索宽度或采样数）来优化小型语言模型文本转SQL性能的方法及效果。
0 citationsRead paper
Selected work
Representative Papers
Cardinal Grid Slime Trail is PSPACE-Complete
This study resolves an open problem concerning the computational complexity of the grid-based Slime Trail puzzle game. Focusing on the cardinal four-directional variant, the authors establish its PSPACE-completeness by presenting a novel reduction from Quantified Boolean Formula (QBF), which introduces specialized gadgets that satisfy both parity and degree constraints inherent to the integer lattice. The construction leverages the regular structure of the grid and naturally extends—via a 45-degree rotation—to the eight-directional version. Consequently, this work proves that both predominant gameplay variants are PSPACE-complete, thereby filling a longstanding gap in the theoretical understanding of this classic puzzle game’s complexity landscape.
0 citationsRead paper
CytoDINO: Risk-Aware and Biologically-Informed Adaptation of DINOv3 for Bone Marrow Cytomorphology
Morphological analysis of bone marrow cells is critical for diagnosing hematologic malignancies, yet manual interpretation is time-consuming, subjective, and entails asymmetric diagnostic risks—misclassifying blasts as normal cells poses significantly greater clinical harm than the reverse. To address this, we propose the first clinical risk-aware, lightweight pathological analysis framework. Our method introduces a Hierarchical Focal Loss (HFL) that incorporates lineage-specific biological priors to impose stronger penalties on high-risk misclassifications. It further integrates parameter-efficient fine-tuning via DINOv3-LoRA, confidence-driven selective prediction, and knowledge-guided modeling of inter-class relationships. Evaluated on the MLL dataset, our framework achieves 88.2% weighted F1 and 76.5% macro-F1 scores while training only 8% of parameters. It is deployable on a single RTX 5080 GPU, and its selective prediction mechanism covers 67% of samples with 99.5% accuracy.
0 citationsRead paper
Recent publications
Latest Papers
Beam Search, Self-Consistency, and the Limits of Inference-Time Scaling for Grammar-Constrained Text-to-SQL in Small Language Models
研究探讨了在语法受限条件下，通过调整推理时计算量（如改变束搜索宽度或采样数）来优化小型语言模型文本转SQL性能的方法及效果。
0 citationsRead paper
Cardinal Grid Slime Trail is PSPACE-Complete
This study resolves an open problem concerning the computational complexity of the grid-based Slime Trail puzzle game. Focusing on the cardinal four-directional variant, the authors establish its PSPACE-completeness by presenting a novel reduction from Quantified Boolean Formula (QBF), which introduces specialized gadgets that satisfy both parity and degree constraints inherent to the integer lattice. The construction leverages the regular structure of the grid and naturally extends—via a 45-degree rotation—to the eight-directional version. Consequently, this work proves that both predominant gameplay variants are PSPACE-complete, thereby filling a longstanding gap in the theoretical understanding of this classic puzzle game’s complexity landscape.
0 citationsRead paper
CytoDINO: Risk-Aware and Biologically-Informed Adaptation of DINOv3 for Bone Marrow Cytomorphology
Morphological analysis of bone marrow cells is critical for diagnosing hematologic malignancies, yet manual interpretation is time-consuming, subjective, and entails asymmetric diagnostic risks—misclassifying blasts as normal cells poses significantly greater clinical harm than the reverse. To address this, we propose the first clinical risk-aware, lightweight pathological analysis framework. Our method introduces a Hierarchical Focal Loss (HFL) that incorporates lineage-specific biological priors to impose stronger penalties on high-risk misclassifications. It further integrates parameter-efficient fine-tuning via DINOv3-LoRA, confidence-driven selective prediction, and knowledge-guided modeling of inter-class relationships. Evaluated on the MLL dataset, our framework achieves 88.2% weighted F1 and 76.5% macro-F1 scores while training only 8% of parameters. It is deployable on a single RTX 5080 GPU, and its selective prediction mechanism covers 67% of samples with 99.5% accuracy.
0 citationsRead paper