IC-386LLM performance on CS-Bench grows logarithmically with parameter scale within model families

Xiaoshuai Song, Muxi Diao, Guanting Dong, Zhengyang Wang, Yujia Fu, Runqi Qiao, Zhexu Wang, Dayuan Fu, Huangxuan Wu, Bin Liang, Weihao Zeng, Yejie Wang, Zhuoma GongQue, Jianing Yu, Qiuna Tan, Weiran Xu

SourceCS-Bench: A Comprehensive Benchmark for Large Language Models towards Computer Science Mastery

The paper evaluates multiple model families (Qwen1.5, Llama2, Llama3, Gemma, InternLM2, DeepSeekLLM) across a range of parameter sizes and finds that CS-Bench scores increase approximately linearly when parameter scale is plotted on a logarithmic axis. For Qwen1.5, the score improves by 16.19% from 0.5B to 7B, 7.11% from 14B to 72B, and only 2.66% from 72B to 110B, showing diminishing returns. A fitting function is validated by predicting Qwen1.5-110b at 67.83% (actual 67.95%) and Llama2-70b at 55.08% (actual 52.52%).

Evidence
correlational
Key metric
Qwen1.5: +16.19% (0.5B→7B), +7.11% (14B→72B), +2.66% (72B→110B); fitting prediction Qwen1.5-110b 67.83% vs actual 67.95%; Llama2-70b predicted 55.08% vs actual 52.52%
Caveat
The fitting function is validated on only two model families (Qwen1.5 and Llama2) and the Llama2 prediction has a 2.56% deviation. The relationship is observed, not causally established.
Model
Qwen1.5, Llama 2 / Llama 2 base, Llama 3, Gemma, InternLM2, DeepSeek LLM
Concepts
Scale-dependent behaviour
Datasets
CS-Bench [eval]
Related findings
IC-387, IC-388, IC-389
Extraction
automatic-extraction