StaCOR: Decision-Stable Output Refinement for TabPFN with Noisy Support Labels

TabPFN-style in-context prediction can be degraded by corrupted support labels. When the deployed support context is cached, version-locked, shared, or otherwise operationally constrained from refresh, external adaptation pipelines can still use newly available clean evidence without modifying that served context. We propose StaCOR (Stability-aware Coupling-Regularized Output Refinement), a selectively constrained fixed-context refinement method for frozen TabPFN. StaCOR leaves the noisy support context unchanged and trains a lightweight output-side correction head using auxiliary clean evidence. Regularized multiclass coupling controls correction capacity at the parameter level, while Selective Decision Stability (SDS) applies a training-time zero-margin penalty only to clean examples that the noisy-support baseline predicts correctly; baseline errors remain unconstrained by this stability term. We evaluate seven tabular datasets, including four development benchmarks and three frozen external-validation datasets under symmetric and instance-dependent support-label noise using probabilistic and decision metrics, paired statistical analysis, direct-context alternatives, and serving measurements. Relative to a regularized affine calibration baseline, StaCOR has a statistically indistinguishable benchmark negative log-likelihood (NLL) under symmetric noise and incurs a small NLL cost under instance-dependent noise, while reducing harmful-flip rates from 0.0195 to 0.0124 and from 0.0202 to 0.0124, respectively. External-validation results show the same decision-stability trend, although the calibration baseline can retain higher Macro-F1. SDS ablation further shows more consistent decision preservation, with small measurable NLL costs in some settings. When context modification is feasible, directly using clean examples generally yields lower NLL and higher Macro-F1. StaCOR therefore targets fixed-context or refresh-constrained deployments where a more favorable probability–decision trade-off is desired without modifying the original served support context.