Dilixiati Muhataer

This PhD project involves developing large foundation models with enhanced capabilities for solving complex, structured tasks such as mathematical reasoning and programming problems. A key objective is to design training and inference techniques that significantly improve model performance and algorithmic efficiency, enabling models to scale to practical reasoning problems with reduced computational cost. The project will also emphasize data curation and filtering, ensuring that models learn from high-quality, diverse, and task-relevant datasets, which are crucial for robust reasoning performance. Beyond improving accuracy, the research will investigate how foundation models can exhibit generalizable algorithmic behavior, allowing them to transfer reasoning skills across domains such as mathematics, logic, and software development. Methodologically, the work will combine insights from optimization, sparsity, and low-rank techniques to make training both more stable and resource-efficient. The student will also explore evaluation frameworks tailored to structured reasoning tasks, which go beyond standard benchmarks and capture the depth and reliability of model reasoning. Ultimately, the project aims not only to push the limits of what large models can achieve in complex reasoning domains, but also to make these advances practically accessible by lowering resource barriers and ensuring reproducibility.