Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
- URL: http://arxiv.org/abs/2510.01624v1
- Date: Thu, 02 Oct 2025 02:57:00 GMT
- Title: Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
- Authors: Feiyang Kang, Michael Kuchnik, Karthik Padthe, Marin Vlastelica, Ruoxi Jia, Carole-Jean Wu, Newsha Ardalani,
- Abstract summary: We study whether high SFT scores translate to improved performance after RL.<n>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.<n>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.
- Score: 20.446287312285648
- License: http://creativecommons.org/licenses/by/4.0/
- 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 $R^2$ 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. Evaluation tool will be open-sourced.
Related papers
- Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models [56.12341509545198]
Large Language Models (LLMs) are pretrained on massive datasets and later instruction-tuned via supervised fine-tuning (SFT) or reinforcement learning (RL)<n>Best practices emphasize large, diverse pretraining data, whereas post-training operates differently.<n>We theoretically analyze transformers trained on an in-context weight prediction task for linear regression.
arXiv Detail & Related papers (2026-03-01T21:58:09Z) - Good SFT Optimizes for SFT, Better SFT Prepares for Reinforcement Learning [8.550698116833123]
Post-training of reasoning LLMs typically consists of an offline SFT stage followed by an online reinforcement learning stage.<n>We show that, after identical RL training, models from stronger SFT checkpoints can significantly underperform those from weaker ones.<n>We propose PEAR, an SFT-stage method that corrects this mismatch and better prepares the model for RL.
arXiv Detail & Related papers (2026-02-01T06:53:45Z) - On the Non-decoupling of Supervised Fine-tuning and Reinforcement Learning in Post-training [10.433802085981046]
Post-training of large language models routinely interleaves supervised fine-tuning (SFT) with reinforcement learning (RL)<n>We show that RL increases SFT loss under SFT optimality and that SFT lowers the reward achieved by RL.<n> Experiments on Qwen3-0.6B confirm the predicted degradation, verifying that SFT and RL cannot be separated without loss of prior performance in the post-training.
arXiv Detail & Related papers (2026-01-12T10:14:09Z) - SWE-RM: Execution-free Feedback For Software Engineering Agents [61.86380395896069]
Execution-based feedback is widely used in the development of coding agents through test-time scaling (TTS) and reinforcement learning (RL)<n>In contrast, execution-free feedback from reward models can provide more fine-grained signals without depending on unit test cases.<n>We introduce SWE-RM, an accurate and robust reward model adopting a mixture-of-experts architecture with 30B total parameters and 3B activated during inference.
arXiv Detail & Related papers (2025-12-26T08:26:18Z) - Trust-Region Adaptive Policy Optimization [82.09255251747818]
Post-training methods play an important role in improving large language models' (LLMs) complex reasoning abilities.<n>We introduce TRAPO, a framework that interleavesSupervised Fine-Tuning (SFT) and Reinforcement Learning (RL) within each training instance.<n>Experiments on five mathematical reasoning benchmarks show that TRAPO consistently surpasses standard SFT, RL, and SFT-then-RL pipelines.
arXiv Detail & Related papers (2025-12-19T14:37:07Z) - Reassessing the Role of Supervised Fine-Tuning: An Empirical Study in VLM Reasoning [30.751908700207185]
SFT plays a crucial role across several scenarios.<n>SFT with only 2K achieves comparable or better reasoning performance to RL with 20K.<n>We identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL.
arXiv Detail & Related papers (2025-12-14T13:46:42Z) - RLSR: Reinforcement Learning with Supervised Reward Outperforms SFT in Instruction Following [4.6740998081727385]
We propose replacing SFT with RLSR to leverage the extensive SFT dataset in an RL framework.<n>In RLSR, the base model generates multiple responses for each prompt, and reward scores are computed as the cosine similarity in the semantic embedding space between the generated and human-labeled responses.
arXiv Detail & Related papers (2025-10-16T01:13:14Z) - RL Fine-Tuning Heals OOD Forgetting in SFT [35.01074051556079]
We investigate the evolution and mechanism behind the synergy ofSupervised Fine-Tuning and Reinforcement Learning.<n>Our findings re-identify the roles of SFT and RL in the two-stage fine-tuning and discover the rotation of singular vectors as the key mechanism.
arXiv Detail & Related papers (2025-09-08T21:40:41Z) - Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved) [3.13388270461847]
We draw on a connection between supervised fine-tuning (SFT) and the theory and practice of finding optimal policies via Reinforcement Learning (RL)<n>We show that a small modification to SFT leads to an importance weighted variant that behaves closer to training with RL as it.<n>We refer to this variant as importance weighted supervised fine-tuning (iw-SFT)
arXiv Detail & Related papers (2025-07-17T07:26:54Z) - Scalpel vs. Hammer: GRPO Amplifies Existing Capabilities, SFT Replaces Them [25.324955028065887]
Two popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT)<n>We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU.<n>SFT exhibits greater updates and also affects mid-layers query more, leading us to hypothesise that this may have caused the out-of-domain degradation.
arXiv Detail & Related papers (2025-07-13T19:04:17Z) - AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy [48.30596996677882]
We investigate the synergy between supervised fine-tuning (SFT) and reinforcement learning (RL) in developing strong reasoning models.<n> scaling strategies yield notable improvements in reasoning performance.<n>Our AceReason-Nemotron-1.1 7B model significantly outperforms AceReason-Nemotron-1.0 and new state-of-the-art performance among Qwen2.5-7B-based reasoning models.
arXiv Detail & Related papers (2025-06-16T09:27:48Z) - Putting the Value Back in RL: Better Test-Time Scaling by Unifying LLM Reasoners With Verifiers [57.95157497749428]
We propose RL$V$ that augments any value-free'' RL method by jointly training the LLM as both a reasoner and a generative verifier.<n> RL$V$ boosts MATH accuracy by over 20% with parallel sampling and enables $8-32times$ efficient test-time compute scaling.
arXiv Detail & Related papers (2025-05-07T22:41:26Z) - OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles [91.88062410741833]
We introduce OpenVLThinker, one of the first open-source large vision-language models (LVLMs) to exhibit sophisticated chain-of-thought reasoning.<n>We show that OpenVLThinker-7B consistently advances performance across six benchmarks demanding mathematical and general reasoning.
arXiv Detail & Related papers (2025-03-21T17:52:43Z) - SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training [127.47044960572659]
Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models.<n>This paper studies the difference between SFT and RL on generalization and memorization.<n>We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants.
arXiv Detail & Related papers (2025-01-28T18:59:44Z) - On the Power of Perturbation under Sampling in Solving Extensive-Form Games [56.013335390600524]
We investigate how perturbation does and does not improve the Follow-the-Regularized-Leader (FTRL) algorithm in solving extensive-form games under sampling.<n>We present a unified framework for textitPerturbed FTRL algorithms and study two variants: PFTRL-KL and PFTRL-RKL.
arXiv Detail & Related papers (2025-01-28T00:29:38Z)
This list is automatically generated from the titles and abstracts of the papers in this site.
This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.