Researcher publishes 'Never Give Up' RL sampling method to counter the 'Matthew Effect' in LLM post-training

A researcher, Michael Noukhovitch, published a blog post with an accompanying arXiv paper and GitHub code introducing "Never Give Up" (NGU), an RL post-training sampling method for LLMs. NGU adaptively retries unsolved hard prompts with probability p before filtering, aiming to fix a "Matthew Effect" where standard RL training (GRPO) mostly improves already-easy problems. Tests on Olmo 3 7B (AIME math), Deepcoder/DeepSWE, and Qwen 2.5 0.5B on GSM8k Platinum reportedly show NGU at k=4 beating fixed-k GRPO across tested values on hard problem subsets.

This is relevant mainly to RL researchers and ML engineers who train or fine-tune LLMs with reinforcement learning, offering a cheap-to-try sampling tweak rather than a new architecture. The practical impact is currently limited to a fairly narrow audience: it does not affect deployed products, end users, or business decisions today. If it holds up, it could inform how labs allocate RL compute during post-training, but that is a technical efficiency gain, not a capability leap.

The "Matthew Effect" framing is a catchy diagnosis but not a new discovery; RL practitioners have long known easy examples dominate gradient signal. There is no independent replication, no evidence of adoption by any lab, and results are confined to small/mid-scale academic benchmarks (a 0.5B and a 7B model), so claims of broad applicability across "math, code, and agentic domains" outrun the evidence shown. The extraction also lacks exact paper and repo URLs, which limits verification.

1. 01Whether the arXiv paper gets cited or replicated by other RL post-training researchers within the next 3-6 months
2. 02Whether any major lab (e.g. those behind Olmo, Qwen, DeepSeek-style RL pipelines) adopts or benchmarks NGU in a production training run
3. 03Whether follow-up experiments extend NGU to larger models (beyond 7B) or to real agentic/code benchmarks like SWE-bench with reported numbers, not just DeepSWE/Deepcoder