The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models
The paper argues that LLM math failures often come from missing the right underlying “primitive,” not from an inability to execute once that structure is supplied.
The authors introduce Mathematical Primitive and a benchmark, HLEI, split across Discovery, Generation, Digestion, and Execution. Their diagnosis says solution accuracy hides different weakness profiles, with Discovery emerging as the main bottleneck. They also claim discovery-limited failures are especially repairable through post-training. Their ABS self-distillation method transfers primitive-guided reasoning into a student model and improves results across model sizes and harder math benchmarks. ArXiv · AI/CL/LG's note
The authors introduce Mathematical Primitive and a benchmark, HLEI, split across Discovery, Generation, Digestion, and Execution. Their diagnosis says solution accuracy hides different weakness profiles, with Discovery emerging as the main bottleneck. They also claim discovery-limited failures are especially repairable through post-training. Their ABS self-distillation method transfers primitive-guided reasoning into a student model and improves results across model sizes and harder math benchmarks. ArXiv · AI/CL/LG's note
score 5