RL Fine-Tuning Heals OOD Forgetting in SFT

• URL: http://arxiv.org/abs/2509.12235v2
• Date: Sat, 01 Nov 2025 13:50:29 GMT
• Title: RL Fine-Tuning Heals OOD Forgetting in SFT
• Authors: Hangzhan Jin, Sitao Luan, Sicheng Lyu, Guillaume Rabusseau, Reihaneh Rabbany, Doina Precup, Mohammad Hamdaqa,
• Abstract summary: 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.
• Score: 35.01074051556079
• License: http://creativecommons.org/publicdomain/zero/1.0/
• Abstract: The two-stage fine-tuning paradigm of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has empirically shown better reasoning performance than one-stage SFT for the post-training of Large Language Models (LLMs). However, the evolution and mechanism behind the synergy of SFT and RL are still under-explored and inconclusive. In our study, we find the well-known claim "SFT memorizes, RL generalizes" is over-simplified, and discover that: (1) OOD performance peaks at the early stage of SFT and then declines (OOD forgetting), the best SFT checkpoint cannot be captured by training/test loss; (2) the subsequent RL stage does not generate fundamentally better OOD capability, instead it plays an \textbf{OOD restoration} role, recovering the lost reasoning ability during SFT; (3) The recovery ability has boundaries, \ie{} \textbf{if SFT trains for too short or too long, RL cannot recover the lost OOD ability;} (4) To uncover the underlying mechanisms behind the forgetting and restoration process, we employ SVD analysis on parameter matrices, manually edit them, and observe their impacts on model performance. Unlike the common belief that the shift of model capacity mainly results from the changes of singular values, we find that they are actually quite stable throughout fine-tuning. Instead, the OOD behavior strongly correlates with the \textbf{rotation of singular vectors}. 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. %reversing the rotations induced by SFT, which shows recovery from forgetting, whereas imposing the SFT parameter directions onto a RL-tuned model results in performance degradation. Code is available at https://github.com/xiaodanguoguo/RL_Heals_SFT

Related papers

• 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)
• 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)
• Rethinking Expert Trajectory Utilization in LLM Post-training [35.018182540417236]
  We propose the Plasticity-Ceiling Framework to ground this landscape.<n>We establish the Sequential SFT-then-RL pipeline as the superior standard.<n>Our findings provide actionable guidelines for maximizing the value extracted from expert trajectories.
  arXiv  Detail & Related papers  (2025-12-12T11:13:00Z)
• Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning [49.22815446849924]
  Large Language Models (LLMs) often struggle with problems that require multi-step reasoning.<n>For small-scale open-source models, Reinforcement Learning with Verifiable Rewards (RLVR) fails when correct solutions are rarely sampled.<n>We propose Supervised Reinforcement Learning (SRL), a framework that reformulates problem solving as generating a sequence of logical "actions"
  arXiv  Detail & Related papers  (2025-10-29T22:05:08Z)
• RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMs [1.772462140285781]
  Training large language models (LLMs) from scratch is increasingly impractical, making post-training methods such as supervised fine-tuning central to modern practice.<n>Using an out-of-distribution (OOD) variant of the 24-point card game and new spectrum-based diagnostics, we revisit how these two stages reshape model representation and OOD performance.
  arXiv  Detail & Related papers  (2025-08-22T17:10:37Z)
• 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)
• Reinforcement Fine-Tuning Enables MLLMs Learning Novel Tasks Stably [80.36077974826865]
  Post-training algorithms such as Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are widely used to adapt multimodal large language models to downstream tasks.<n>We study the behavior of SFT and RFT on an open-source multimodal model, Qwen2.5-VL.<n>Our experiments reveal a sharp trade-off: SFT enables rapid task acquisition but leads to catastrophic forgetting, whereas RFT learns more slowly on novel tasks but maintains prior knowledge.
  arXiv  Detail & Related papers  (2025-06-30T04:15:01Z)
• 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)
• Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning [93.00629872970364]
  Reinforcement learning (RL) has become the dominant paradigm for improving the performance of language models on complex reasoning tasks.<n>We introduce SPARKLE, a fine-grained analytic framework to dissect the effects of RL across three key dimensions.<n>We study whether difficult problems -- those yielding no RL signals and mixed-quality reasoning traces -- can still be effectively used for training.
  arXiv  Detail & Related papers  (2025-06-05T07:53:59Z)
• How Much Backtracking is Enough? Exploring the Interplay of SFT and RL in Enhancing LLM Reasoning [6.92510069380188]
  We investigate the dynamics between SFT and RL on eight reasoning tasks.<n>Short CoT sequences used in SFT as a warm-up do have moderate contribution to RL training, compared with cold-start RL.<n>We find that longer CoT with backtracks generally induce better and more stable RL training.
  arXiv  Detail & Related papers  (2025-05-30T06:49:00Z)
• 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.