GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training
For practitioners of large language model post-training, GAC provides a lightweight method to dynamically balance supervised fine-tuning and reinforcement learning signals, improving performance especially at larger model scales.
GAC introduces a noise-aware adaptive mixing controller for hybrid SFT-RL post-training that adjusts the mixing weight based on gradient variance and signal disagreement, achieving consistent improvements over fixed and rule-based baselines across math, code, science, and logic benchmarks with less than 1% training overhead.
Hybrid post-training usually combines supervised fine-tuning and reinforcement learning, but fixed mixing schedules cannot adapt when the relative noise of the two signals changes over time. We propose GAC, a noise-aware controller that derives an adaptive mixing weight from online estimates of gradient variance and disagreement between the two training signals. The method adds smoothing, prior guidance, and bounded updates while reusing existing training tensors. Experiments on math, code, science, and logic benchmarks show that GAC consistently improves hybrid post-training over strong fixed and rule-based baselines, with larger gains at larger model scales and less than 1% training overhead.