What matters in AI.

Subscribe

Research

MagicBeanStalk increases the score of gpt-oss-120b from 61.2 to 69.8

About 88% of the problems in a large math corpus have the same result in all rollouts and give no learning signal.

Claimed, not confirmed

Roughly 88% of problems give no learning signal A ten by ten grid: 88 red squares are problems solved or failed in every rollout; 12 outlined squares are the rest. Training on the full corpus gained only 1.1 points. ZERO-ADVANTAGE TRAP roughly 88% of problems Solved or failed every rollout The rest Training on the full corpus: +1.1 points
In one large math corpus, roughly 88% of problems were solved or failed in every attempt, so training on all of it gained just 1.1 points.

Sources

  1. MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesisarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·12 OCT2026

Posted

Tags

More in Research

All Research news