Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension Ability

TL;DR

The paper tackles whether large language models truly understand deep semantics or rely on surface cues. It builds a causal mediation framework that defines direct causal effect ($DCE$) for deep structure and indirect causal effect ($ICE$) for surface structure, introducing estimable surrogates $ADCE$ and $AICE$ and an intervention-based data generation protocol. Empirically, it shows widespread deep-structure comprehension across mainstream LLMs, with deep-structure reliance increasing with accuracy and scale, and reveals differences between closed- and open-source models. The proposed $ADCE$/$AICE$ methodology offers a bidirectional, more robust assessment than accuracy alone, especially under spurious correlations, providing a principled tool for evaluating deep semantic understanding in LLMs.

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, 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 rely more on deep structure, while open-source LLMs 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.

Figures (13)

• Figure 1: Surface structure interventions cause subtle accuracy degradation relative to the obvious accuracy decline from deep structure changes.
• Figure 2: Approximated DCE (ADCE) quantifies LLMs' deep structure comprehension, while approximated ICE (AICE) measures surface structure understanding. Comparing them reveals LLMs' reliance on deep or surface structures. Our method involves: initial inference, intervention on correct samples, and secondary inference for ADCE and AICE calculation. More details are in \ref{['alg: alg adce']}.
• Figure 3: Causal graph with mediation: ${\bm{x}} \to d \to Y$ shows deep structures' direct causal effect, ${\bm{x}} \to s \to Y$ indicates surface structures' indirect causal effect via mediator $s$.
• Figure 4: For the four intervention strategies, LLM accuracy drops from 100% when surface structures are altered while deep structures remain unchanged in initially correct samples.
• Figure 5: Deep structure understanding in LLMs via ADCE. Positive ADCE demonstrate the existence of direct causal effect of deep structure on outcomes, increasing with model scale and accuracy. Accuracy-DCE slopes vary across tasks, with steeper slopes indicating higher task complexity and greater reliance on various deep structure comprehension ability.

Theorems & Definitions (3)

• Theorem 1
• Theorem 2
• proof