Fine-tuning LLMs on 30K-90K AgenticMath samples matches or surpasses baselines trained on 400K-2.3M samples on math reasoning benchmarks¶

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

Subclaims¶

q2 i? AgenticMath-Llama3-8B fine-tuned on 90K samples achieves an average score of 40.3, surpassing Llama3-8B-DART-Math trained on 590K samples (39.7) and other large-scale Llama3 baselines. → 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? · causal · r2

Benchmark evaluation (Table 2) comparing AgenticMath fine-tuned Llama3-8B at 30K/60K/90K samples against large-scale baselines (400K-2.3M) across six math benchmarks; the text reports "AgenticMath-Llama3-8B (90K) achieves 40.3" versus DART-Math's 39.7 at 590K.

"AgenticMath-Llama3-8B (90K) achieves 40.3, surpassing Llama3-8B-DART-Math (590K, 39.7) and all other large-scale Llama3 baselines."

Discussion¶

Related Claims¶

• Increasing AgenticMath-generated training data from 7.5K to 22.5K samples yields consistent accuracy gains for Llama3-8B across benchmarks — related
• At 30K and 60K data scales, AgenticMath outperforms all compared baselines across four base models on six math benchmarks — related