ResearchMath-14K: Scaling Research-Level Mathematics via Agents

ResearchMath-14K scales research-level mathematics by converting open problems from the literature into agent-generated training data: the authors curate 1,233 academic sources from arXiv, zbMATH, workshops, and repositories, then use a two-stage extractor-refiner pipeline to produce 14,056 self-contained problems, the largest public research-grade math corpus to date. They also release ResearchMath-Reasoning, 220K teacher trajectories from GPT-OSS-120B and Qwen3-30B-A3B, and show that trace quality is fragile: in 720 sampled traces, 54.0% contain at least one fake reference, while newer model generations emit 5.6× more reference-like mentions and 5.0× more fabricated references per trace. A GPT-5.5 self-containment audit raises extracted questions from 67.2% to 94.2%, and duplicate filtering uses Qwen3-Embedding-8B with a 0.9 similarity threshold. Most importantly, LoRA fine-tuning Qwen3-4B, Qwen3-8B, and Qwen3-30B-A3B on a 5,000-trace filtered subset improves downstream performance by 9.2 percentage points on average, outperforming a DASD-Thinking control in 8 of 9 benchmark cells and indicating that wrong-but-reasonable research attempts can still provide useful supervision after filtering hallucinated citations and non-attempts.