AI Research
New Coding Benchmark: 5.4% Passed, Not The Quoted 47/100
The paper's central claim is that behavioural tests let agents pass by copying the original implementation. It names the failure mode, then measures how often it happens.
5d ago
#technical debt