IC-396Closed-source LLMs (GPT, Claude) rely more on deep structure than open-source LLMs (Llama, Mistral), and open-source models' surface sensitivity decreases with model scale

Yujin Han, Lei Xu, Sirui Chen, Difan Zou, Chaochao Lu

SourceBeyond Surface Structure: A Causal Assessment of LLMs' Comprehension ability

By comparing ADCE (deep structure effect) and AICE (surface structure effect) for 12 released LLMs across five tasks, the paper finds that closed-source models (GPT-3.5-Turbo, GPT-4o, Claude-3-Sonnet, Claude-3.5-Sonnet) show a larger ADCE-minus-AICE gap, indicating greater reliance on core semantics. Open-source models (Llama-2, Llama-3, Mistral, Mixtral) show higher AICE, meaning their outputs are more sensitive to surface-level presentation changes. Within open-source families, this surface sensitivity diminishes as model size increases (e.g., Llama-2-7B to Llama-2-70B, Llama-3-8B to Llama-3-70B, Mixtral-8x7B to Mixtral-8x22B).

Evidence
correlational
Caveat
The comparison is based on the ADCE-AICE difference shown in Figure 7; specific per-model values are not printed in the text. The finding is limited to the four model families tested and five tasks evaluated.
Model
Llama 2 / Llama 2 base Llama 2 7B, Llama 2 13B, Llama 2 70B, Llama 3 8B, 70B, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1, Mixtral 8x7B / Mistral 8x7B Instruct / Mixtral 46.7B / Mixtral 8x7B Instruct / Mixtral-instruct-8x7b, Mixtral 8x22B, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, GPT-4o, Claude 3 Sonnet, Claude 3.5 Sonnet
Concepts
Scale-dependent behaviour, Shortcut
Datasets
GSM8K [eval], Analytic Entailment [eval], CommonsenseQA [eval]
Related work
Hooda et al. 2024 [context], Guo et al. 2024 [context]
Related findings
IC-395
Extraction
automatic-extraction