Xamt Cross-Framework Differential Fuzzing Finds 72 Deep-Learning Library Discrepancies That CPU-vs-GPU Oracles Miss

arXiv 2608.11886·medium signalXBlueskyLinkedInCopy link

Most DL library API testing relies on intra-library properties or CPU–GPU differential oracles, which are blind to defects that behave consistently across hardware backends. Xamt instead constructs execution-validated groups of APIs meant to implement equivalent operations across seven libraries, using explicit aliases and parameter-role normalization, then explores them with variance-guided differential fuzzing over ordinary, boundary, and non-finite inputs. It built 676 validated groups covering 2,563 matched APIs and found 72 independently reproduced discrepancies — 4 crashes and 68 output inconsistencies — of which 25 were confirmed and 23 fixed.