IC-395Most mainstream LLMs exhibit positive ADCE across five tasks, indicating reliance on deep structure for problem-solving, with ADCE strongly correlated with accuracy (r² > 0.7)

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

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

The paper measures the Approximated Direct Causal Effect (ADCE) of deep structure on model outputs for 12 released LLMs across five tasks (2-digit multiplication, GSM8K, word unscrambling, analytic entailment, CommonsenseQA). ADCE is positive for most models, in contrast to the zero ADCE of a random-weights baseline, indicating that intervening on core semantics causes previously correct answers to become incorrect. A strong linear correlation exists between accuracy and ADCE across all tasks, with r² values ranging from 0.784 to 0.977, suggesting that higher-accuracy models depend more on deep structure.

Evidence
correlational
Key metric
r² = 0.859 (2-digit multiplication, y=0.54x+0.43), r² = 0.784 (GSM8K, y=0.81x+0.19), r² = 0.846 (word unscrambling, y=0.60x+0.13), r² = 0.862 (analytic entailment, y=1.64x-0.61), r² = 0.977 (CommonsenseQA, y=0.93x+0.03)
Caveat
ADCE is an approximation of the true direct causal effect; the approximation relies on intervention strategies that minimize but do not eliminate the gap between the true ICE and AICE. Some model-task pairs (e.g., Llama-3-8B on analytic entailment) show negative ADCE.
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
Datasets
GSM8K [eval], Analytic Entailment [eval], CommonsenseQA [eval]
Methods
Causal mediation analysis / Vig et al. 2020 (causal mediation analysis) [primary]
Related work
Stolfo et al. 2022 [builds-on], Gonzalez & Nori 2024 [context], Jiang et al. 2024b [context]
Related findings
IC-396
Extraction
automatic-extraction