IC-030Larger language models exhibit higher independence rates and lower conformity under some protocols

Zhiyuan Weng, Guikun Chen, Wenguan Wang

SourceDo as We Do, Not as You Think: the Conformity of Large Language Models

The paper finds that model size correlates positively with independence rates across model families. For example, Qwen2 series IR rises from 19.6% at 7B to 57.6% at 72B; Llama3.1-405B has the second-highest IR at 56.1%. Conversely, smaller models show higher conformity, with Qwen2-7B showing 98.7% CRC and 95.3% CRW. The paper also notes architectural and training differences affect conformity, e.g., Llama3.1-70B has low CRW of 9.2% attributed to larger high-quality datasets and long-context pretraining.

Evidence
correlational
Key metric
Independence rates: Qwen2-7B 19.6%, Qwen2-72B 57.6%; Llama3.1-405B 56.1%. Conformity rates: Qwen2-7B CRC=98.7%, CRW=95.3%, CRT=79.4%, CRD=27.5%; Llama3.1-70B CRC=12.0%, CRW=9.2%, CRT=15.6%, CRD=73.5%.
Caveat
The precise causes of differences across model sizes and families remain unclear due to limited public information on training processes and alignment strategies.
Model
Qwen 2, Llama 3, Llama 3.1, Gemma 2, GPT-4o, GPT-3.5 / ChatGPT-3.5
Concepts
Scale-dependent behaviour
Datasets
BENCHFORM [eval]
Related work
Meta Llama 3.1 technical report [context], Qwen technical report [context]
Extraction
automatic-extraction