When Can You Correct Distribution Drift in Temporal Graph Generation? A Sharpening--Drift Tension and an Impossibility for Observation-Based Correction
The paper argues that temporal graph drift creates an error floor that past observations cannot reliably correct.
Li and coauthors derive the degradation from the masked flow-matching loss, tying it to structures that are rare in training but common at deployment. Their experiments find drift-period marginal error barely changes across a 50x sampling-budget range, while the drift error floor sits 2.2x to 34.3x above the in-period floor. The claimed correction limit is formal: any observation-based corrector leaves at least the conditional variance of the tracked statistic. In their tests, an oracle removes 60% of the error, while the best observation-based corrector recovers only 5.7% of that. ArXiv · AI/CL/LG's note
Li and coauthors derive the degradation from the masked flow-matching loss, tying it to structures that are rare in training but common at deployment. Their experiments find drift-period marginal error barely changes across a 50x sampling-budget range, while the drift error floor sits 2.2x to 34.3x above the in-period floor. The claimed correction limit is formal: any observation-based corrector leaves at least the conditional variance of the tracked statistic. In their tests, an oracle removes 60% of the error, while the best observation-based corrector recovers only 5.7% of that. ArXiv · AI/CL/LG's note
score 4