跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2601.07389 · v3

The Post-Training Dilemma: Why We Should Rethink the Sequential SFT-RL Paradigm

首次提交 2026-01-12 18:14
版本更新 2026-11-19 09:42
雷达观测 2026-11-20 10:50
arXiv ↗PDF ↗

摘要

This position paper argues that the sequential supervised fine-tuning followed by reinforcement learning (SFT-then-RL) pipeline---the de facto paradigm for large language model (LLM) post-training---has a fundamental property that cannot be fixed by hyperparameter tuning: SFT and RL optimize over nearly orthogonal parameter subspaces, so RL inevitably degrades the capabilities acquired during the preceding SFT phase. Under the Polyak-Lojasiewicz (PL) condition, which is empirically observed for over-parameterized networks, we show that RL increases the SFT loss by at least μ2dist(θRL,ΘSFT∗)2\fracμ{2}\mathrm{dist}(θ_{\mathrm{RL}}, Θ^*_{\mathrm{SFT}})^2 for some μ>0μ>0. Although the KL penalty can constrain distributional shift, it cannot correct the underlying gradient misalignment between the two objectives. Experiments on Qwen3-0.6B across four tasks confirm the theoretical predictions: on MATH-500, Alpaca, HH-RLHF, and SST-2, the SFT loss increases after the RL phase. We identify the underlying mechanism as persistent near-orthogonality between the gradients of the two objectives, paired with opposing landscape dynamics: SFT flattens the loss landscape while RL re-sharpens it. Based on our analysis, we propose three promising research directions: (1) joint optimization of the two objectives with gradient surgery to resolve conflicting updates; (2) landscape-aware scheduling that terminates RL training at the theoretically optimal point; and (3) characterizing the model scale threshold above which SFT becomes unnecessary.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v3,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v3 · 2026-11-19 09:42 · 最新观测版本

The Post-Training Dilemma: Why We Should Rethink the Sequential SFT-RL Paradigm

This position paper argues that the sequential supervised fine-tuning followed by reinforcement learning (SFT-then-RL) pipeline---the de facto paradigm for large language model (LLM) post-training---has a fundamental property that cannot be fixed by hyperparameter tuning: SFT and RL optimize over nearly orthogonal parameter subspaces, so RL inevitably degrades the capabilities acquired during the preceding SFT phase. Under the Polyak-Lojasiewicz (PL) condition, which is empirically observed for over-parameterized networks, we show that RL increases the SFT loss by at least μ2dist(θRL,ΘSFT∗)2\fracμ{2}\mathrm{dist}(θ_{\mathrm{RL}}, Θ^*_{\mathrm{SFT}})^2 for some μ>0μ>0. Although the KL penalty can constrain distributional shift, it cannot correct the underlying gradient misalignment between the two objectives. Experiments on Qwen3-0.6B across four tasks confirm the theoretical predictions: on MATH-500, Alpaca, HH-RLHF, and SST-2, the SFT loss increases after the RL phase. We identify the underlying mechanism as persistent near-orthogonality between the gradients of the two objectives, paired with opposing landscape dynamics: SFT flattens the loss landscape while RL re-sharpens it. Based on our analysis, we propose three promising research directions: (1) joint optimization of the two objectives with gradient surgery to resolve conflicting updates; (2) landscape-aware scheduling that terminates RL training at the theoretically optimal point; and (3) characterizing the model scale threshold above which SFT becomes unnecessary.

查看此版本解读 →查看此版本原文 ↗