Evaluating invariance training with certified answer-changing controls in neural graph algorithms

Impact 5 arXiv cs.AI

Summary

  • Researchers examine whether invariance training evaluation detects lost sensitivity to answer-changing edits in neural graph algorithms using minimum spanning trees.
  • Using certified controls on 16-vertex and 64-vertex graphs, tests evaluate how graph transformers and message-passing networks respond to structural modifications.
  • Invariance objectives lower violations but can reduce sensitivity to answer-changing attachments, a trade-off avoided by processing biconnected blocks separately.

Why it matters

Evaluating robustness in neural algorithmic reasoners helps ensure that improvements in invariance do not inadvertently degrade model sensitivity to critical input changes.

Who should care