Sensitivity of ReCAST to Teacher Model Choice

Determine how sensitive ReCAST’s robustness gains are to the choice of teacher model used for training-data construction and structured supervision.

Background

ReCAST relies on a single teacher model, DeepSeek-V4-Pro, to generate synthetic training data, mine hard examples, and produce restoration-oriented annotations. The authors note that this dependence may introduce correlated biases in the types, styles, and coverage of obfuscation patterns represented in the training corpus, potentially limiting generalization to naturally occurring, ambiguous, or evolving obfuscations. Consequently, the effect of selecting a different teacher model on ReCAST’s robustness remains unresolved.

References

This may affect the representativeness of the resulting training corpus, particularly for naturally occurring, ambiguous, or evolving obfuscation patterns, so the robustness gains may not fully generalize, and the sensitivity of ReCAST to teacher choice remains an open question.

— ReCAST: Restoration-aware Cascaded Stage-wise Training for Obfuscated SMS Risk Classification  (2609.04878 - Huang et al., 4 Sep 2026) in Limitations, second limitation