LLM-Generated Backends Show Statistically Significant Memory Growth Over 48-Hour Runs in JavaScript, Python and Rust
arXiv 2608.26391 generated service-based applications from BaxBench backend scenarios in three languages, validated them with BaxBench-derived tests, then subjected them to 48-hour workload executions while monitoring memory, response time and throughput, analysed with the Mann-Kendall test and Sen's slope estimator. Memory usage is the most consistent aging indicator, with statistically significant upward trends in most application-language combinations, while response time and throughput behave heterogeneously. Static analysis identified plausible code-level aging mechanisms, and human-written implementations showed aging too, so the takeaway is that passing functional tests says nothing about whether a generated service survives continuous operation.
↳ Follow the thread