Learned Look-Ahead Splitting Rule for CART
提出了一种前瞻性的CART树构建方法，通过预测误差减少评估每个候选分裂点，并使用节点特征学习下游分裂值以提高计算效率。
0 citationsRead paper
Selected work
Representative Papers
Evidence of Phase Transitions in Small Transformer-Based Language Models
This work investigates whether phase transitions—akin to those observed in large language models—emerge during the training of small Transformer language models, and whether they can be directly observed early in training on a linear time scale. To address this, the study employs character-level GPT-style models and introduces high-sensitivity dynamical probes: mean token length, fraction of correctly predicted tokens, and lexical diversity—monitored via Poisson and sub-Poisson statistical analyses to circumvent the insensitivity of conventional loss curves. Results demonstrate that sharp, unambiguous phase transitions occur robustly in the early training stage of small models, without requiring logarithmic time rescaling, and exhibit cross-scale universality. This constitutes the first empirical confirmation that phase transitions are an intrinsic, scale-invariant feature of language model training. The proposed probe metrics establish a novel analytical paradigm for studying training dynamics, enabling fine-grained, real-time characterization of emergent linguistic structure.
0 citationsRead paper
Project Patti: Why can You Solve Diabolical Puzzles on one Sudoku Website but not Easy Puzzles on another Sudoku Website?
This study addresses the inconsistency in difficulty ratings across Sudoku websites. We propose two novel, unsupervised, and quantifiable difficulty metrics: (1) a structural complexity measure based on clause-length distribution derived from SAT encoding; and (2) a simulation-based solver integrating four human-like solving strategies with randomized Nishio backtracking. Together, these form a cross-platform difficulty standardization framework. Evaluated on over 1,000 puzzles from five major Sudoku websites, our approach achieves strong agreement with original site labels—Spearman’s ρ > 0.85 on four sites. It successfully establishes a universal three-tier classification (Easy/Medium/Hard) and supports novice-oriented solving guidance. To our knowledge, this is the first fully automated, interpretable, and platform-agnostic Sudoku difficulty assessment method that requires no human annotation.
0 citationsRead paper
Recent publications
Latest Papers
Learned Look-Ahead Splitting Rule for CART
提出了一种前瞻性的CART树构建方法，通过预测误差减少评估每个候选分裂点，并使用节点特征学习下游分裂值以提高计算效率。
0 citationsRead paper
Evidence of Phase Transitions in Small Transformer-Based Language Models
This work investigates whether phase transitions—akin to those observed in large language models—emerge during the training of small Transformer language models, and whether they can be directly observed early in training on a linear time scale. To address this, the study employs character-level GPT-style models and introduces high-sensitivity dynamical probes: mean token length, fraction of correctly predicted tokens, and lexical diversity—monitored via Poisson and sub-Poisson statistical analyses to circumvent the insensitivity of conventional loss curves. Results demonstrate that sharp, unambiguous phase transitions occur robustly in the early training stage of small models, without requiring logarithmic time rescaling, and exhibit cross-scale universality. This constitutes the first empirical confirmation that phase transitions are an intrinsic, scale-invariant feature of language model training. The proposed probe metrics establish a novel analytical paradigm for studying training dynamics, enabling fine-grained, real-time characterization of emergent linguistic structure.
0 citationsRead paper
Project Patti: Why can You Solve Diabolical Puzzles on one Sudoku Website but not Easy Puzzles on another Sudoku Website?
This study addresses the inconsistency in difficulty ratings across Sudoku websites. We propose two novel, unsupervised, and quantifiable difficulty metrics: (1) a structural complexity measure based on clause-length distribution derived from SAT encoding; and (2) a simulation-based solver integrating four human-like solving strategies with randomized Nishio backtracking. Together, these form a cross-platform difficulty standardization framework. Evaluated on over 1,000 puzzles from five major Sudoku websites, our approach achieves strong agreement with original site labels—Spearman’s ρ > 0.85 on four sites. It successfully establishes a universal three-tier classification (Easy/Medium/Hard) and supports novice-oriented solving guidance. To our knowledge, this is the first fully automated, interpretable, and platform-agnostic Sudoku difficulty assessment method that requires no human annotation.
0 citationsRead paper