New Optimizer PsiLogic Enhances Adam for Stable AI Model Training

Standard adaptive optimizers like Adam and AdamW apply a consistent update rule throughout the entire training process, regardless of whether the model is in an unstable initial phase or nearing convergence. Researchers have introduced PsiLogic, an enhanced optimizer that augments Adam with a dynamic "Active Cancellation Term." This term is controlled by a dual exponential moving average of scale-normalized gradient norms, effectively acting as a "chaos detector." When gradient statistics are unstable, PsiLogic increases damping, and as training stabilizes, this damping gradually diminishes, providing an automatic warmup without requiring manual scheduling. PsiLogic was rigorously evaluated against Adam, AdamW, and Lion using "FairBench," a new reproducible benchmark protocol. This protocol ensures fair comparisons through per-optimizer learning-rate sweeps, identical initialization across seeds, and statistical Welch t-tests. On an NVIDIA H100, PsiLogic demonstrated superior validation metrics in three out of four tested domains: NLP perplexity, ViT top-1 accuracy, and ResNet top-1 accuracy, showing statistically significant improvements over existing optimizers. While diffusion model performance was comparable, and some ResNet comparisons were numerically tied, the overall results highlight PsiLogic's effectiveness. The authors have released an open-source PyTorch implementation, the FairBench harness, and all raw data for independent verification.