Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead

Feiyang Kang, Michael Kuchnik, Karthik Padthe, Marin Vlastelica, Ruoxi Jia, Carole-Jean Wu, Newsha Ardalani

Keywords: Post-Training, Large Reasoning Models, Large Language Models, Performance Prediction, Reinforcement Learning with Verifiable Rewards

TL;DR: We show extensive examples where high SFT scores do not transfer to improved RL performance in reasoning post-training; we propose generalization loss on held-out SFT examples and pass@large k as viable proxies for predicting post-RL performance.

Abstract:

In post-training for reasoning Large Language Models (LLMs), the current state of practice trains LLMs in two independent stages: Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR, shortened as "RL" below). In this work, we challenge whether high SFT scores translate to improved performance after RL. We provide extensive counter-examples where this is not true. We find high SFT scores can be biased toward simpler or more homogeneous data and are not reliably predictive of subsequent RL gains or scaled-up post-training effectiveness. In some cases, RL training on models with improved SFT performance could lead to substantially worse outcome compared to RL on the base model without SFT. We study alternative metrics and identify generalization loss on held-out reasoning examples and Pass@large k performance to provide strong proxies for the RL outcome. We trained hundreds of models up to 12B-parameter with SFT and RLVR via GRPO and ran extensive evaluations on 7 math benchmarks with up to 256 repetitions, spending
1M GPU hours. Experiments include models from Llama3, Mistral-Nemo, Qwen3 and multiple state-of-the-art SFT/RL datasets. Compared to directly predicting from pre-RL performance, prediction based on generalization loss and Pass@large k achieves substantial higher precision, improving
coefficient and Spearman's rank correlation coefficient by up to 0.5 (2x). This provides strong utility for broad use cases. For example, in most experiments, we find SFT training on unique examples for a one epoch underperforms training on half examples for two epochs, either after SFT or SFT-then-RL; With the same SFT budget, training only on short examples may lead to better SFT performance, though, it often leads to worse outcome after RL compared to training on examples with varying lengths. This work develops an enhanced evaluation tool that will be open-sourced.

Submission Number: 32