A new paper on arXiv examines a common technique for probing whether language models memorize their training data: inserting synthetic records, or “canaries,” that contain known secrets. If the model later reveals those secrets, researchers take it as evidence of memorization. The paper asks a more basic question: how much do these measurements depend on the canaries themselves?

The study focuses on supervised fine-tuning (SFT) and the role of masked context in controlling memorization. The abstract frames the central issue as a potential blind spot—if canary-based results shift with the format, placement, or masking of the canary, then they may not provide a stable measure of memorization.

Because the abstract is truncated, the full findings are not available in the source. Still, the paper’s significance is clear: it challenges the assumption that canary tests are neutral instruments, and it points toward a need for more careful interpretation of memorization benchmarks in fine-tuned models.