Under the same teacher model, AgenticMath outperforms prior data-generation methods, indicating gains arise from framework design rather than teacher strength¶

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? In a 10K-sample controlled comparison using the same GPT-4o-mini teacher, AgenticMath achieves the highest average performance (31.6) versus MetaMath (26.9), DARTMath (29.9), and ScaleQuest (26.3). → 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

Controlled 10K-sample comparison (Table 5) across MetaMath, DARTMath, ScaleQuest, and AgenticMath, all trained under the same SFT configuration on Mistral-7B and using the same teacher model for solution generation; the text reports "AgenticMath achieves the highest av- erage performance among all methods."

"As shown in Table 5, AgenticMath achieves the highest av- erage performance among all methods. This con- trolled experiment confirms that the observed im- provements do not arise from using a stronger teacher model"

Discussion¶

Related Claims¶

• Each AgenticMath component contributes performance gains, with the review-revise loop playing a particularly important role — related
• At 30K and 60K data scales, AgenticMath outperforms all compared baselines across four base models on six math benchmarks — related
• Increasing AgenticMath-generated training data from 7.5K to 22.5K samples yields consistent accuracy gains for Llama3-8B across benchmarks — related
• Review-revise iterations improve synthesized problem quality, with gains largely stabilizing by the third iteration — related