IC-698GPT-3.5-turbo's CoT reasoning errors are correlated across different demonstration sets, while PoT errors are less correlated

Murong Yue, Jie Zhao, Min Zhang, Liang Du, Ziyu Yao

SourceLarge Language Model Cascades with Mixture of Thought Representations for Cost-Efficient Reasoning

When GPT-3.5-turbo fails on a hard reasoning question under chain-of-thought prompting, re-prompting it with a different set of CoT demonstrations typically yields the same incorrect answer. In contrast, program-of-thought prompting tends to produce a different (or correct) answer. Among cases where CoT fails, 43% are solved by PoT and 57% receive a different wrong answer from PoT. This error-correlation structure is what makes mixing CoT and PoT (MOT) more effective at detecting model uncertainty than either representation alone.

Evidence
correlational
Key metric
43% of CoT-failure cases solved by PoT, 57% different wrong answer from PoT (Appendix D case analysis)
Caveat
The case analysis is illustrative and the 43%/57% split is reported without a formal sample size or confidence interval; the pattern is demonstrated on a subset of hard questions.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo
Concepts
Failure mode
Datasets
GSM8K [eval], Big-Bench Hard [eval], CREPE [eval]
Related work
FrugalGPT [compared-to]
Related findings
IC-699, IC-700
Extraction
automatic-extraction