New benchmark challenges AI coding tests, shows 64.2% failure rate in real-world scenarios

Researchers have introduced TFD-Bench, a new benchmark designed to evaluate AI models in multi-turn software debugging scenarios. The benchmark addresses limitations in current AI coding assessments, which typically test single-turn code generation rather than iterative development processes. It reveals that state-of-the-art language models fail 64.2% of real-world regression tests despite generating syntactically correct code. The study found that incorporating test execution feedback significantly improves issue resolution rates while reducing computational waste. TFD-Bench evaluates models across 50 tasks covering common Python error categories through a five-phase closed-loop testing cycle.