Designing Robust Learning as Phase Control: Identical Training Produces Stable Success or Irreversible Failure

Learning from noisy labels is a central challenge in machine learning, yet existingapproaches assume performance degrades smoothly with increasing noise.Here we show that this assumption can fail dramatically. Under dual-gradientlearning—where training mixes gradients from noisy labels and a small trustedsubset—the mixing coefficient λ induces sharp transitions between qualitativelydistinct outcomes. Across 3,916 runs in vision and language benchmarks,identical conditions produce either stable success or irreversible failure, withbimodality, hysteresis gaps of 22–53 percentage points, and an absorbing trapstate (0/18 escape attempts). Gradient geometry distinguishes regimes andtracks degradation. Cross-domain validation reveals shared sharp boundariesalongside domain-specific signatures. These findings reframe robust learningas phase control—mapping hazardous regions, monitoring gradient alignment,and designing protocols that avoid irreversible failure—with broad implicationsfor systems where noisy supervision or fine-tuning trajectories permanentlydetermine outcomes.