GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training

In three linesGAC is an adaptive controller for hybrid SFT-RL post-training that dynamically adjusts mixing weights based on online estimates of gradient variance and disagreement between the two training signals. Tested on math, code, science, and logic benchmarks, GAC improves fixed baselines with less than 1% computational overhead.

Summary generated by Claude — human-verified