Alibaba tops AIME25 with agentic data synthesis

The real bottleneck for reasoning model training has always been data quality, not model scale — and this paper makes the case more forcefully than most with a surprisingly small 11K trajectory count.

• –The core insight: treat problem synthesis as a multi-step agentic task (Draft → Check → Refine → Finalize) with a library of composable atomic skills, so the proposer can build genuinely hard, verifiable problems rather than trivially easy or unsolvable ones
• –Multi-Granularity Policy Optimization (MGPO) is the RL secret sauce — it combines trajectory-level and stage-level advantage estimates to handle sparse rewards during proposer training, outperforming standard GRPO by 6.5 points
• –Cross-domain generalization is notable: gains aren't just on math (AIME, HMMT) but extend to coding (LiveCodeBench +5pts), science (OlympicArena +4.4%), and general reasoning (GPQA +6.3%)
• –The 91.6% AIME25 figure needs scrutiny — no independent replication yet, and the GitHub repo is an empty placeholder; the results hinge on their proprietary verifier ensemble (gpt-oss-120b, DeepSeek-V3.2-Special, Qwen3-235B-Thinking)
• –If the code/weights release materializes, this could meaningfully shift how labs approach cold-start synthetic data pipelines for reasoning models