This blog post discusses a paper on RL post-training of LLMs, introducing the Matthew Effect and proposing 'Never Give Up' to solve it. The author presents interactively, with deeper technical details in the paper and code.

Every RL practitioner has seen an eval curve go up, like AIME 2025 during training of Olmo 3.1 RL-Zero Math. But what does it mean? Breaking 30 AIME questions into difficulty levels based on initial pass@32, we find easy problems improve drastically, while hard problems (initial pass@32=0) barely improve. Averaging hid this: gains come from easier problems. This is the Matthew Effect.

We see similar in code RL and agentic RL using Deepcoder and Deep SWE. Gains are proportional to initial competence. We connect this to the Matthew Effect in network science and economics: 'the rich get richer.' The Matthew Effect in RL for LLMs: RL improves performance in proportion to initial competence—easy tasks get easier, hard tasks remain hard.

Causes? Not just GRPO. If no correct answer in k samples, no gradient. One fix: sample more completions (larger k). We test with Qwen 2.5 0.5B Instruct on GSM8k platinum, varying k.