IC-389All evaluated LLMs score significantly lower on CS reasoning questions than knowledge questions, with the gap narrowing for stronger models

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

Across all 30+ evaluated models, the average score on reasoning-type CS questions is 45.60% compared to 60.61% on knowledge-type questions, a gap of approximately 15 percentage points. The paper observes that weaker models show a much larger relative gap (e.g., Llama2-7b scores 26.19% on open-ended reasoning but 60.61% on assertion reasoning), while stronger models like GPT-4 score comparably across formats (68.94% open-ended reasoning vs 67.68% assertion reasoning). Error analysis attributes the primary cause of failures to lack of domain knowledge rather than general reasoning deficiency, and notes that CS-specific reasoning errors (logical reasoning) do not decrease as much as mathematical reasoning errors with model scale.

Evidence
correlational
Key metric
Average reasoning 45.60% vs knowledge 60.61%; Llama2-7b: open-ended reasoning 26.19% vs assertion reasoning 60.61%; GPT-4: open-ended reasoning 68.94% vs assertion reasoning 67.68%
Caveat
The gap is measured on a single benchmark (CS-Bench) and may not generalize to other domains. The error type analysis is based on GPT-4 categorization of MC question errors, which may introduce classification bias.
Model
Llama 2 / Llama 2 base, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-4o, GPT-3.5 / ChatGPT-3.5, PaLM 2, Claude 2.1
Concepts
Scale-dependent behaviour
Datasets
CS-Bench [eval]
Related findings
IC-386, IC-387, IC-388
Extraction
automatic-extraction