Original title: From Rollouts to Recipes: Self-Contained Post-Training for LLMs
Yifei Li, Lingling Zhang, Muye Huang, Zihan Ma +2 · September 2026
Training large language models usually means treating every practice problem the exact same way. However, a model's trial answers, called rolloutsThe model's generated outputs or attempts when solving a problem, used to evaluate its performance during training., show that it understands some problems much better than others. Researchers built a method called Self-Routing to fix this. It checks how accurate and confident the model is on each generated attempt. Then, it picks the right training step for that specific item. A sample might get special reinforcement training, learn from its own good answers, undergo stabilizing tweaks, or get skipped completely. This system needs no outside teachers or extra labels. In math tests, Self-Routing consistently beat standard training methods. It shifted its choices as learning progressed, successfully cutting out wasted effort on useless or already mastered problems.
Imagine teaching a cooking class where students have different skill levels. A beginner needs hands-on guidance for every step, while an expert might only need feedback on advanced techniques. In this paper, the training system does something similar: it assesses how well the model is learning from each example and chooses the best way to teach it—providing detailed help, refining skills, adding safeguards, or skipping if the model already has it mastered.
This helps AI developers build more efficient models by focusing training effort where it's needed, potentially improving performance in tasks like mathematical reasoning without extra human work or annotations.
The abstract only reports results for mathematical reasoning with Qwen3 and Qwen3.5 model backbones, so it's unclear how well this approach generalizes to other tasks or different language models. It doesn't discuss computational costs or implementation details.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
From Rollouts to Recipes: Self-Contained Post-Training for LLMs — Yifei Li, Lingling Zhang, Muye Huang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.