Research map / Large language models / Language, safety and interpretability
Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension ability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes approximated direct causal effect (ADCE) and approximated indirect causal effect (AICE) via causal mediation analysis to measure whether LLMs rely on deep structure (core semantics) versus surface structure, and compares closed vs open-source LLMs.
Problem
Whether LLMs truly comprehend deep structure (core semantics) or rely mainly on surface structure; existing evaluations emphasize surface sensitivity and accuracy, which can miss deep-structure reliance.
Approach
Formulates deep-structure comprehension as a direct causal effect and surface-structure reliance as an indirect effect within a parallel multiple-mediator causal model, then estimates them with observable approximations (ADCE, AICE) based on mask/rephrase interventions on correctly answered samples, using indicator outcomes; shows ADCE equals a weighted sum of probability of sufficiency and necess
Key results
- Most tested LLMs show positive ADCE that grows with accuracy (R^2 > 0.7 linear fit), while the random-weight Llama-3-70b has zero ADCE and accuracy.
- Closed-source models (GPT, Claude) rely more on deep structure than surface structure, while open-source Llama/Mistral are more surface-sensitive, with the sensitivity decreasing as model scale increases.
- On CivilComments with increasing spurious correlation, accuracy in the majority group stays high while ADCE declines, showing ADCE better reflects reliance on core semantics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase model scale within the same model series | Llama-2, Llama-3, and Mixtral models · Five tasks | ADCE | Helped |
| Increase prediction accuracy | Evaluated LLMs · Five tasks | ADCE: R^2 > 0.7 | Helped |
| Apply supervised fine-tuning to activate task-specific knowledge | Llama-3-8B · Analytic Entailment | ADCE | Helped |
| Use larger model scale | Open-source LLMs, including Llama · Evaluated tasks | AICE | Hurt |
| Increase spurious-correlation levels in training data | Llama-3 · CivilComments majority group | ADCE | Hurt |
| Increase spurious-correlation levels in training data | Llama-3 · CivilComments majority group | accuracy | No clear effect |
| Increase spurious-correlation levels in training data | Llama-3 · CivilComments minority group | accuracy | Mixed |
| Increase the deep-structure weight c1 | Synthetic logistic-regression model · Synthetic data | ADCE | Helped |
| Increase the deep-structure weight c1 | Synthetic logistic-regression model · Synthetic data | AICE | Hurt |
Surprising findings
- Llama-3-8b shows negative ADCE on Analytic Entailment, meaning altering deep structure can improve accuracy.
- SFT on Analytic Entailment raised Llama-3-8b's ADCE, suggesting deep-structure comprehension depends on activating task knowledge.
Limitations
- ADCE and AICE are approximations of the non-estimable original DCE/ICE, relying on the similarity of intervention data.
- Rephrasing relies on Claude-3.5-Sonnet as an agent with a self-checking mechanism, which may introduce errors.
- Evaluation restricted to five tasks and binary indicator outcomes.
Experimental setup
Datasets: 2-digit multiplication GSM8k Word Unscrambling Analytic Entailment CommonsenseQA CivilComments
Benchmarks: 2-digit multiplication GSM8k Word Unscrambling Analytic Entailment CommonsenseQA
Metrics: ADCE AICE accuracy
Baselines: accuracy random-weight Llama-3-70b
Builds on: causal mediation analysis probability of necessity and sufficiency
Closest papers by meaning
- Enhancing LLM’s Cognition via Structurization (NeurIPS 2024 · similarity 0.89)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.88)
- Do Language Models Use Their Depth Efficiently? (NeurIPS 2025 · similarity 0.87)
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.87)
- Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.