AI & ML
The Post-Training Dilemma: Why We Should Rethink the Sequential SFT-RL Paradigm
arXiv math.IT0 views
arXiv:2601.07389v3 Announce Type: replace-cross Abstract: 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 $\frac{\mu}{2}\mathrm{dist}(\theta_{\mathrm{RL}}, \Theta^*_{\mathrm{SFT}})^2$ for some $\mu>0$. 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.
Source:https://arxiv.org/abs/2601.07389Related posts
Trying to get our Arms around AI efficiency
0Stack Overflow Blog
Why an Embedded Game Fails on Some Phones, and How to Debug It
0dev.to
Microsoft 365 vs Office 2024: The Break-Even Math
0dev.to
Master AI Evals: Build Reliable, Scalable Model Evaluation Pipelines
0dev.to
What to give someone who misses a place
0dev.to
Google is pushing OEMs against GrapheneOS and bootloader unlocking
0Hacker News