MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis

The paper introduces MagicBeanStalk, a recursive data‑synthesis method that transforms seed math problems into new, policy‑ and budget‑calibrated training problems. By avoiding the zero‑advantage trap found in large open‑source RLVR corpora, the approach yields significant performance gains on multiple math benchmarks and financial datasets. Experiments show that continual RLVR with MagicBeanStalk problems boosts GPT‑OSS‑120B from 61.2 to 69.8 and DeepSeek v4Flash from 72.4 to 77.9 on a 16K‑token budget, outperforming training on the full source corpus.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms
• benchmarks