IC-700GPT-4's reasoning accuracy degrades when provided with incorrect hints from a weaker model

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

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

In the mot-1d-verify cascade, when the weaker LLM (GPT-3.5-turbo) produces two inconsistent answers, those answers can be passed as hints to GPT-4 before it answers. On the subset of questions routed to GPT-4, adding these hints slightly improves accuracy on Date (0.892 to 0.910) but significantly hurts performance on GSM8K (0.891 to 0.867) and CREPE (0.774 to 0.727). The authors attribute the degradation to GPT-4 being misled by the incorrect answers when both hints are wrong.

Evidence
correlational
Key metric
GSM8K: 0.891 (w/o hints) vs 0.867 (w/ hints); CREPE: 0.774 vs 0.727; Date: 0.892 vs 0.910
Caveat
The experiment is conducted only on the subset of questions that the mot-1d-verify approach routes to the stronger LLM, not on the full dataset.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Failure mode
Datasets
GSM8K [eval], Big-Bench Hard [eval], CREPE [eval]
Methods
Self-Consistency / Self-consistency prompting / Wang et al. 2023 (self-consistency) / Wang et al. 2023b (Self-Consistency) [primary]
Related work
Progressive-hint prompting [builds-on]
Related findings
IC-698, IC-699
Extraction
automatic-extraction