Lottery Tickets Are Not Deployment Tickets
Accuracy-matched sparse models still changed deployment behavior enough to break drop-in replacement.
The paper tests whether lottery tickets or other sparse challengers can replace a dense incumbent without changing fixed downstream decision logic. Across its experiments, sparse models often recovered clean accuracy but still differed on calibration, OOD behavior, class reliability, representations, and policy decisions. In small-gap fixed-threshold settings, lottery-ticket replacement changed 7% to 10% of accept-review decisions. The paper argues that clean accuracy is not a sufficient deployment certificate, because small confidence shifts near an operating boundary can create routing churn. ArXiv · AI/CL/LG's note
The paper tests whether lottery tickets or other sparse challengers can replace a dense incumbent without changing fixed downstream decision logic. Across its experiments, sparse models often recovered clean accuracy but still differed on calibration, OOD behavior, class reliability, representations, and policy decisions. In small-gap fixed-threshold settings, lottery-ticket replacement changed 7% to 10% of accept-review decisions. The paper argues that clean accuracy is not a sufficient deployment certificate, because small confidence shifts near an operating boundary can create routing churn. ArXiv · AI/CL/LG's note
score 4