MagicBeanStalk lifts gpt-oss-120b from 61.2 to 69.8 points on math

The authors of the MagicBeanStalk preprint argue that most open RLVR data is useless for any given model. In a combined math corpus of 78K tasks, about 88% of problems, for a fixed policy and rollout budget, are either solved on every attempt or failed on every attempt. The proposed algorithm recursively grows new tasks from hard ones, calibrated by difficulty. As a result, gpt-oss-120b rose from 61.2 to 69.8 points (+8.6) on eight math benchmarks, and DeepSeek v4Flash from 72.4 to 77.9 (+5.4), with a budget of 16K tokens.

The zero-advantage trap

RLVR, reinforcement learning with verifiable rewards, is usually scaled by adding more tasks. The authors suggest looking at something else: whether a dataset gives a useful signal to a specific model at a specific budget. They describe dataset effectiveness as a function of three variables: the tasks themselves, the target policy, and the number of rollouts per training step.

In methods with group-normalized rewards, such as GRPO, if all attempts on a task get the same result, the advantage is zero and there is no gradient. According to the paper, almost the entire combined corpus of popular RLVR datasets falls into exactly this situation. Easy tasks are always solved, hard ones never are, and there is nothing to learn from.

Filtering helps partly: it concentrates the remaining tasks that show a spread of outcomes. But it only selects, and in doing so it discards the hardest examples, which is exactly where the reasoning the model needs lies. A simple workaround fails too: asking a teacher LLM to vary hard tasks directly does not produce the right calibration.

How the algorithm works

MagicBeanStalk works in two passes: top-down decomposition and bottom-up recursive synthesis. From a source task (the seed), it extracts reasoning operations and the ways they are composed. It then assembles families of fundamentally new tasks from them, preserving the original's "reasoning blueprint." Difficulty is tuned to the policy and the budget: some rollouts should succeed and some should fail, and at least one correct solution must fit within the training limit.

Key figures from the paper:

• gpt-oss-120b: 61.2 → 69.8 (+8.6) on 8 math benchmarks, 16K-token budget;
• DeepSeek v4Flash: 72.4 → 77.9 (+5.4) under the same conditions;
• training on the full source corpus of 78K tasks: only +1.1;
• best filtered subset of real data: +5.8;
• finance: more than 14K verifiable tasks synthesized from just 340 solutions;
• gain over the base model in finance: +6.9 on SecQue and +2.4 on FinanceBench.

The comparison with filtering is telling: synthetic data beats the best real-data slice by 2.8 points for gpt-oss-120b. Meanwhile, the sheer size of the 78K-task corpus turns out to be nearly useless.

What teams training models should take away

The practical takeaway is simple: before launching RLVR, measure the share of tasks with zero outcome variance for your model and your rollout budget. If it is close to 88%, a significant part of the compute will be wasted, and it makes more sense to calibrate the dataset than to enlarge it. The finance example shows that the method can be run even where labeled data is scarce: 340 solutions were enough for 14K verifiable tasks.

The caveats are obvious. This is a first-version preprint, and the results have not been peer-reviewed and were obtained by the authors themselves. The gain in finance is more modest than in math, and the method requires verifiable answers, so it cannot be applied in domains without a clear verifier. Independent reproductions on other models and budgets will show how robust the conclusions are.