Never Give Up: a fix for the Matthew Effect in RL, where training keeps improving problems the model already solves
Michael Noukhovitch published NGU on September 15 (submitted to HN by Nathan Lambert), naming the failure mode as the Matthew Effect: standard GRPO concentrates compute on easy problems while hard ones stagnate, and a scalar average hides it. NGU starts with small completion counts and re-queues unsolved problems with probability p, producing a geometric distribution of attempts weighted toward the hard cases. On GSM8k, k=4 with p=0.9 beat every standard GRPO configuration on the hardest subsets; on DeepScaler with Qwen 3 4B base (~120 H100 hours) it improved AIME 2025 and BRUMO 2025 without regressing easy problems. The transferable lesson for anyone running RL or evals is to read fine-grained per-difficulty metrics rather than a single average.
↳ Follow the thread