BuggyMLE: Scaling Machine Learning Engineering Difficulty with Bug Injection
4.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2611.25088.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryBuggyMLE introduces a repository-level machine-learning engineering benchmark, a Code Complexity Index, over 6,600 bug-injected variants, and a difficulty-graded benchmark. A fine-tuned Qwen3.5-9B matches frontier models on BuggyMLE-Bench-Lite and improves results on SWE-bench Lite and LiveCodeBench.
Why it mattersThe structural difficulty metric and graded bug corpus make agent evaluation more realistic and provide a reusable basis for comparing repository-scale MLE performance.
Cited by
No citations on record.