TLDR Tech / Dev
Learning to Solve Hard Problems in RL for LLMs by Never Giving Up
Standard RL post-training disproportionately improves problems the base model already solves, a pattern the piece calls the Matthew Effect. Never Give Up counters it by continuing to sample hard examples, with experiments across math and code showing better gains on previously unsolved tasks while introducing tradeoffs around asynchronous staleness and compute.
Read the full story