MagicBeanStalk: 88% of RLVR math corpus is a zero-advantage trap, fix via recursive synthesis
stanfordnlp · x · 2026-10-15
arXiv paper by Ximing Lu, Aviral Kumar, Jan Kautz, Yejin Choi, and David Acuna shows a large open-source RLVR math corpus is almost entirely a zero-advantage trap: 88% of problems are solved in every rollout or failed in every rollout for a fixed policy and rollout budget, giving no learning signal. Filtering can only select, not create hard problems.
Their fix, MagicBeanStalk, is a top-down decomposition + bottom-up recursive data-synthesis algorithm that grows seed problems into families of new, policy- and budget-calibrated training problems — hard enough to require the reasoning compositions of difficult problems, but calibrated so some rollouts succeed within budget.
More from Research
- Initialization matters far more than expected in LLM-driven scientific discovery, new study finds — BenBlaiszik · 2026-10-15
- RLVR is just RL — RLHF is the weird one, argues lateinteraction — lateinteraction · 2026-10-15
- Mathematician models OpenAI's 372-theorem drop as a self-organized criticality collapse — littmath · 2026-10-15
- RoPE long-context failures proven inevitable; training-free RoPE Profiler boosts accuracy up to 25% — zhaofeng_wu · 2026-10-15
- OpenAI's Quasi-Riemann proof sparks chain of results, Bhargava proves primitive root theorem — Singularitarian · 2026-10-15
- Applied Compute's Spork post-training recipe doubles chess puzzle solve-rate vs GRPO — ypatil125 · 2026-10-15