IC-1309GPT-4-0613 exhibits reduced output diversity relative to GPT-3.5: it outperforms on best@1 but underperforms on best@8 under CoT prompting

Alexander G Shypula, Aman Madaan, Yimeng Zeng, Uri Alon, Jacob R. Gardner, Yiming Yang, Milad Hashemi, Graham Neubig, Parthasarathy Ranganathan, Osbert Bastani, Amir Yazdanbakhsh

SourceLearning Performance-Improving Code Edits

Under chain-of-thought prompting, GPT-4-0613 achieves 26.99% optimization rate at best@1 versus 21.37% for GPT-3.5, but at best@8 GPT-4 reaches only 42.74% opt / 1.58x speedup while GPT-3.5 reaches 43.05% opt / 1.60x. The authors attribute this to a lack of output diversity in GPT-4 despite using the same sampling hyper-parameters (temperature 0.7). The gap is small in absolute terms but consistent across the CoT condition.

Evidence
correlational
Key metric
GPT-4 CoT: best@1 26.99% opt, 1.32x; best@8 42.74% opt, 1.58x. GPT-3.5 CoT: best@1 21.37% opt, 1.25x; best@8 43.05% opt, 1.60x.
Caveat
The authors use the hedged language 'this may demonstrate a lack of output diversity'; the best@8 gap is small (0.31% opt, 0.02x speedup) and could reflect sampling variance rather than a systematic diversity difference.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5
Concepts
Failure mode
Methods
Chain-of-Thought prompting / Chain-of-Thought (CoT) prompting / CoT prompting / Few-shot Chain of Thought / Few-shot CoT / Chain-of-Thought (CoT-S) / CoT-bag prompting / Few-shot CoT prompting / Rationale prompting / Zero-shot CoT prompting / Wei et al. (2022) chain-of-thought prompting [primary]
Related findings
IC-1307, IC-1308, IC-1310
Extraction
automatic-extraction