MagicBeanStalk: 88% of RLVR data trapped in zero advantage, recursive synthesis adds 8.6 points
Lianhuiq · x · 2026-10-15
An arXiv paper by Ximing Lu, Aviral Kumar, Jan Kautz, Yejin Choi, and David Acuna shows 88% of pooled open-source RLVR math problems fall into a "zero-advantage trap": for a fixed policy and rollout budget they pass or fail every rollout, yielding no learning signal, and filtering merely discards hard problems. MagicBeanStalk synthesizes policy- and budget-calibrated problems via top-down decomposition plus bottom-up recursive synthesis, guided by three empirical principles (hard seeds specify what to synthesize; preserve reasoning operations; calibrate difficulty to policy and budget). Results: +8.6 points across 8 math benchmarks for gpt-oss-120b and +5.4 for DeepSeek v4 Flash.
More from Research
- New paper builds a categorical theory unifying hundreds of PEFT methods, proving LoRA rank limits — aryaman2020 · 2026-10-15
- ActionOPD distills agents at action level, lifting Qwen3 3.3-6.9 points on OOD benchmarks — nics-efc · 2026-10-15
- Horizon Robotics' VITAL hits 90.8 PDMS with a single denoising step at 3.29x inference speed — HorizonRobotics · 2026-10-15
- KAIST's KV4-RL: INT4 KV cache quantization speeds RLVR rollouts 1.8x without accuracy loss — kaist-ai · 2026-10-15
- ASU's τ-Emotion-Bench shows LLM agents stay unreliable when users get emotional — Arizona-State-University · 2026-10-15
- Huawei's SkillMarker watermarks agent skills, detecting copied passages of 1,536 tokens — Alan Tang · 2026-10-15