Higher seed-problem filtering thresholds yield better downstream reasoning performance, indicating problem quality directly impacts results¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Using Mistral-7B, raising the Stage 1 filtering threshold from score 2 to 3 to 4 improves average benchmark accuracy (33.4, 34.9, 35.0). → Xianyang Liu 2025

Evidence¶

Xianyang Liu 2025¶

Xianyang Liu, Yilin Liu, Shuai Wang, Hao Cheng, Andrew Estornell, Yuzhi Zhao, Jun Shu, Jiaheng Wei. (2025). AgenticMath: Enhancing LLM Reasoning via Agentic-based Math Data Generation. arXiv preprint. https://arxiv.org/abs/2510.19361 doi:10.48550/arxiv.2510.19361

q2 · i? · design · r2

Threshold sensitivity experiment with Mistral-7B at a fixed 30K dataset (Table 4): score thresholds of 2, 3, and 4 yield average scores of 33.4, 34.9, and 35.0 respectively, though the main experiments adopt threshold 3.

"Table 4 shows that the use of higher thresholds for the filtering of seed problems leads to better results, confirming that the quality of the selected problems directly impacts the performance of reasoning."

Discussion¶

Related Claims¶

• Each AgenticMath component contributes performance gains, with the review-revise loop playing a particularly important role — related
• Raising the answer-confidence threshold reduces the number of retained MCQs but substantially improves their quality — related
• Instances occurred where the Teacher LLM directly provided the final answer, bypassing the intended cognitive process, and can be programmatically filtered — related
• Seeding student groups with alternative ideas improves learning more than randomly assigning students to groups — related