Support-Contra Asymmetry in LLM Explanations

1. Problem Statement

Large Language Models (LLMs) increasingly generate natural language explanations (rationales) alongside their predictions to aid interpretability. However, a critical gap exists in understanding whether these explanations genuinely reference the predictive cues present in the input text or if they are merely plausible-sounding post-hoc narratives.

• The Challenge: Unlike traditional machine learning models where feature importance is quantifiable, LLM explanations are free-form text. They may appear fluent and convincing even when they do not correspond to measurable evidence driving the prediction.
• The Core Question: Do LLM-generated explanations accurately track the predictive lexical evidence in the input, and does this alignment differ between correct and incorrect predictions?

2. Methodology

The authors propose an empirical framework to measure explanation–evidence alignment by comparing LLM-generated text against predictive signals extracted from transparent, external reference models.

A. Experimental Setup

• Datasets: Three benchmark text classification datasets were used:
  • WIKIONTOLOGY: Multi-class entity classification (15 classes).
  • AG NEWS: Four-class news topic classification.
  • IMDB: Binary sentiment analysis (Positive/Negative).
• LLM Generation: Predictions and explanations were generated using DeepSeek-R1 with deterministic decoding (temperature=0).
• Reference Models: To establish a ground truth for "predictive evidence," the authors trained transparent linear classifiers on the same datasets:
  • Logistic Regression (LogReg)
  • Linear Support Vector Machines (SVM)
  • Feature Representation: TF-IDF (unigrams and bigrams).

B. Feature Extraction & Partitioning

For each input instance, the reference model identifies the top-$k$ most influential lexical features ($k \in \{3, 5, 7\}$). These features are partitioned based on their contribution to the predicted label:

1. Supporting Features ($S_i$): Features with positive weights/contributions that increase the likelihood of the predicted class.
2. Contradicting Features ($C_i$): Features with negative weights/contributions that provide evidence against the predicted class.

C. Alignment Metrics

The study measures how often these features appear in the LLM's explanation ($e_i$) using two matching strategies:

• Token-aware matching: Normalized matching (lowercasing, lemmatization, punctuation removal).
• Exact matching: Strict string equality.

Key Metrics:

• Support Coverage: Proportion of supporting features found in the explanation.
• Contra Coverage: Proportion of contradicting features found in the explanation.
• Asymmetry Metric ($\Delta^*$): Defined as $\Delta_{support} - \Delta_{contra}$, where $\Delta$ represents the difference in mean coverage between correct predictions (Match) and incorrect predictions (Mismatch).

3. Key Contributions

1. Identification of Support–Contra Asymmetry: The paper empirically characterizes a consistent pattern where explanations for correct predictions emphasize supporting evidence and suppress contradicting evidence, while explanations for incorrect predictions do the opposite.
2. Quantitative Framework: Introduction of a simple, reproducible method to measure explanation–evidence alignment by comparing free-text explanations against signed lexical evidence from transparent classifiers.
3. Robustness Validation: Demonstration that this asymmetry holds across different datasets, reference model families (LogReg vs. SVM), and feature retrieval depths ($k$).
4. Behavioral Insight: Evidence suggesting that LLM explanations are not purely unconstrained narratives but often reflect the distribution of predictive lexical signals present in the input, particularly when the model is correct.

4. Key Results

The study observed a consistent Support–Contra Asymmetry across all experimental conditions:

• Correct Predictions (Match): Explanations showed higher coverage of supporting features and lower coverage of contradicting features.
• Incorrect Predictions (Mismatch): Explanations showed significantly higher coverage of contradicting features (cues that oppose the predicted label) compared to correct predictions.

Statistical Findings:

• WIKIONTOLOGY: Showed the strongest asymmetry. Correct predictions had ~0.52 support coverage vs. ~0.30 for mismatches; contradiction coverage was ~0.12 for matches vs. ~0.21 for mismatches.
• AG NEWS & IMDB: Followed the same qualitative trend, though the magnitude of the asymmetry was slightly lower in IMDB (likely due to the more diffuse nature of sentiment signals compared to topic/entity classification).
• Robustness: The pattern persisted regardless of whether LogReg or SVM was used as the reference, and remained stable across $k=3, 5, 7$.
• Feature Ablation: Removing "supporting" features from input text caused more correct predictions to fail, while removing "contradicting" features improved accuracy, confirming these cues are behaviorally relevant to the LLM.

5. Significance and Implications

• Diagnostic Value: The asymmetry serves as a diagnostic tool. If an explanation references high levels of contradicting evidence, it is a strong indicator that the underlying prediction is likely incorrect.
• Faithfulness vs. Plausibility: The results challenge the notion that LLM explanations are purely post-hoc rationalizations. Instead, they suggest that when an LLM is correct, its explanation tracks the actual predictive signals in the text. When it is wrong, the explanation often rationalizes the error by focusing on misleading cues.
• Evaluation Paradigm: The paper advocates for evaluating explanations not just on human plausibility, but on their alignment with external, interpretable evidence.
• Limitations: The study relies on linear classifiers (TF-IDF), which may miss higher-order semantic or compositional patterns that LLMs utilize. Additionally, lexical matching cannot capture semantic paraphrasing.

Conclusion

The paper establishes that LLM explanations exhibit a robust support–contra asymmetry. Correct predictions are accompanied by explanations that align with supporting lexical evidence, whereas incorrect predictions are associated with explanations that inadvertently highlight contradicting cues. This finding provides a new, empirical lens for evaluating the reliability and faithfulness of LLM-generated rationales in text classification tasks.