IC-898GPT-3.5-turbo and GPT-4 are overconfident in their initial code predictions when unit test execution is unavailable

Xinyun Chen, Maxwell Lin, Nathanael Schärli, Denny Zhou

SourceTeaching Large Language Models to Self-Debug

When self-debugging is performed without unit test execution, the model must rely solely on self-generated feedback to assess correctness. GPT-3.5-turbo shows no improvement on Transcoder (89.1 to 89.1) and only 0.6% on MBPP (67.6 to 68.2). GPT-4 shows 0.9% on Transcoder (77.3 to 78.2) and 3.6% on MBPP (72.8 to 76.0). The authors explicitly state that both models 'tend to be overconfident in their initial predictions' in this setting, in contrast to codex which still improves by up to 5% without execution.

Evidence
correlational
Key metric
Transcoder without execution: GPT-3.5 89.1→89.1, GPT-4 77.3→78.2; MBPP without execution: GPT-3.5 67.6→68.2, GPT-4 72.8→76.0
Caveat
The overconfidence is observed specifically in the self-debugging prompting context; it is unclear whether it generalizes to other self-correction or self-evaluation settings.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Failure mode
Datasets
Transcoder [eval], MBPP [eval]
Related findings
IC-897
Extraction
automatic-extraction