GoKawiilResearchers identify 'Matthew Effect' limiting RL training gains on hard math problems in LLMs
A new paper examines RL post-training of the Olmo 3 model on AIME math problems and finds that reported accuracy gains mask an uneven pattern: easy problems improve dramatically while the hardest problems, which the base model initially fails entirely, barely improve at all. The authors call this the 'Matthew Effect' and propose a technique called 'Never Give Up' to address it.
GoKawiil's interpretation of the reporting above, not reported fact.
This exposes a blind spot in how RL training progress for LLMs is typically measured, since aggregate eval scores can create a false impression of broad reasoning improvement when gains are really concentrated on problems the model could already partially solve. If RL training systematically fails to teach genuinely hard problem-solving, it raises questions about whether current post-training methods actually push model capability boundaries or just polish existing skills.
- Averaged eval scores can obscure whether RL training improves genuinely hard problems or just easy ones.
- On AIME math problems, Olmo 3 showed strong gains only on problems it could already partially solve before RL.
- The 'Never Give Up' method is proposed to counter this Matthew Effect and target unsolved hard problems.
Source: mnoukhov.github.io — Sep, 2026-09-15
Published there as: “Learning to solve hard problems in RL for LLMs by never giving up”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.