Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension ability
Abstract
Large language models (LLMs) have shown remarkable capability in natural language tasks, yet debate persists on whether they truly comprehend deep structure (i.e., core semantics) or merely rely on surface structure (e.g., presentation format). Prior studies observe that LLMs' performance declines when intervening on surface structure, arguing their success relies on surface structure recognition. However, surface structure sensitivity does not prevent deep structure comprehension. Rigorously evaluating LLMs' capability requires analyzing both, yet deep structure is often overlooked. To this end, we assess LLMs' comprehension ability using causal mediation analysis, aiming to fully discover the capability of using both deep and surface structures. Specifically, we formulate the comprehension of deep structure as direct causal effect (DCE) and that of surface structure as indirect causal effect (ICE), respectively. To address the non-estimability of original DCE and ICE -stemming from the infeasibility of isolating mutual influences of deep and surface structures, we develop the corresponding quantifiable surrogates, including approximated DCE (ADCE) and approximated ICE (AICE). We further apply the ADCE to evaluate a series of mainstream LLMs (and the one with random weights), showing that most of them exhibit deep structure comprehension ability, which grows along with the prediction accuracy. Comparing ADCE and AICE demonstrates closed-source LLMs (e.g., GPT) rely more on deep structure, while open-source LLMs (e.g., Llama) are more surface-sensitive, which decreases with model scale. Theoretically, ADCE is a bidirectional evaluation, which measures both the sufficiency and necessity of deep structure changes in causing output variations, thus offering a more comprehensive assessment than accuracy, a common evaluation in LLMs. Our work provides new insights into LLMs' deep structure comprehension and offers novel methods for LLMs evaluation. The code for our project is available at https://github.com/OpenCausaLab/ADCE.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
Cited by top-tier papers1
Builds on18
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah et al.NeurIPS 2020 · 64,255 citations
- LoRA: Low-Rank Adaptation of Large Language ModelsEdward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu et al.ICLR 2022 · 18,833 citations
- Finetuned Language Models are Zero-Shot LearnersJason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu et al.ICLR 2022 · 4,966 citations
- Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought PromptingMiles Turpin, Julian Michael, Ethan Perez, Samuel R. BowmanNeurIPS 2023 · 1,792 citations
- WILDS: A Benchmark of in-the-Wild Distribution ShiftsPang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie et al.ICML 2021 · 1,773 citations
Related papers
- Measuring Intent Comprehension in LLMsNadav Kunievsky, James EvansICML 2026 · 1 citation
- Causality-Aided Evaluation and Explanation of Large Language Model-Based Code GenerationZhenlan Ji, Pingchuan Ma, Zongjie Li, Zhaoyu Wang et al.ISSTA 2025 · 1 citation
- Towards a Holistic and Automated Evaluation Framework for Multi-Level Comprehension of LLMs in Book-Length ContextsYuho Lee, Jiaqi Deng, Nicole Hee-Yeon Kim, Hyangsuk Min et al.EMNLP 2025
- Can Large Language Models Understand Intermediate Representations in Compilers?Hailong Jiang, Jianfeng Zhu, Yao Wan, Bo Fang et al.ICML 2025
- Challenging Large Language Models with New Tasks: A Study on their Adaptability and RobustnessChenxi Li, Yuanhe Tian, Zhaxi Zerong, Yan Song et al.ACL Findings 2024 · 3 citations
