https://blog.lukesalamone.com/tags/grpo/ Recent content in Grpo on Luke Salamone's Blog Hugo en-us Sat, 26 Sep 2026 02:36:55 -0700 https://blog.lukesalamone.com/posts/matthew-effect/ Sat, 26 Sep 2026 02:36:55 -0700 https://blog.lukesalamone.com/posts/matthew-effect/ <p><a href="https://arxiv.org/pdf/2609.13443v1">Learning to Solve Hard Problems in RL for LLMs by Never Giving Up</a> discusses an approach for countering what the authors call the <em>Matthew Effect</em>, the tendency for the LLM to improve much more on easy problems that it is already good than harder problems that have a lower solve rate. Their approach is to allocate training time dynamically based on the difficulty of the problem.</p> <h2 id="background">Background</h2> <p>After an LLM has been pretrained on a large corpus of supervised fine-tuning (SFT) data, it is common to post-train <a href="../../posts/notes-on-deepseek-r1/">using reinforcement learning methods like GRPO</a>. This allows the model to improve on tasks with verifiable rewards and even exceed the performance of the original SFT data.</p>