The authors analyzed binary-reward RLVR, where model answers receive a verifiably correct or incorrect score.
Recursive synthesis turns unsolved problems into useful reinforcement learning data
MagicBeanStalk decomposes problems that are too hard for a model, then recombines their reasoning steps into calibrated training tasks.
Big Tech
Ximing Lu · Aviral Kumar · Jan Kautz · Yejin Choi · David Acuna
University of Washington · Carnegie Mellon University · NVIDIA · Stanford University
Research Digest··3 min read
Lu et al.
Why this paper
From NVIDIA and 3 others
In one line
Most problems in a large RLVR math corpus are too easy or too hard for a fixed policy; effective synthetic data grows from hard seeds with calibrated difficulty.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (3 noted)
- ✓Reports numbers on named benchmarks (5 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§