IC-853Multi-agent debate with GPT-3.5-turbo-0301 does not outperform self-consistency at equivalent inference cost on GSM8K

Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou

SourceLarge Language Models Cannot Self-Correct Reasoning Yet

The paper replicates the multi-agent debate method of Du et al. (2023) using three GPT-3.5-turbo-0301 agents over two rounds on the full GSM8K test set. At 6 responses, multi-agent debate scores 83.2 versus 85.3 for self-consistency; at 9 responses, 83.0 versus 88.2. The authors conclude that the improvement attributed to 'debate' is actually due to self-consistency (sampling and voting), not to the critique mechanism.

Evidence
correlational
Key metric
GSM8K: standard prompting (1 response) 76.7; self-consistency (3) 82.5; multi-agent debate (6) 83.2; self-consistency (6) 85.3; multi-agent debate (9) 83.0; self-consistency (9) 88.2
Caveat
Only one model (GPT-3.5-turbo-0301) and one dataset (GSM8K) are tested for this comparison. The authors use the exact prompt from Du et al. (2023) with 3 agents and 2 rounds.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo
Datasets
GSM8K [eval]
Methods
Multi-Agent Debate [primary], Self-Consistency / Self-consistency prompting / Wang et al. 2023 (self-consistency) / Wang et al. 2023b (Self-Consistency) [compared-to]
Related work
Multi-Agent Debate [builds-on]
Related findings
IC-852, IC-854
Extraction
automatic-extraction