Can Large Language Models Still Explain Themselves? Investigating the Impact of Quantization on Self-Explanations

AI summary

Overview

Research area: Natural Language Processing, specifically explainable AI (XAI) for large language models and efficient model deployment via quantization.

Technical level: Intermediate. Readers need some familiarity with quantization concepts (post-training quantization, weight-only quantization, bit widths) and with explanation evaluation metrics, but the paper explains its setup in accessible terms.

Scope: A comprehensive empirical and human-study investigation of how three post-training quantization techniques at different bit widths affect the quality and faithfulness of two types of self-explanations generated by six LLMs ranging from 7B to 72B parameters.

What This Paper Is About

Self-explanations (SEs) are statements LLMs generate to justify their own outputs, and they are increasingly used to make black-box models transparent in high-stakes settings. Quantization — reducing parameter precision to compress models and speed up inference — is widely deployed, but whether it degrades a model's ability to explain itself had not been characterized. This paper systematically tests how quantization affects the quality and faithfulness of two free-text SE types: natural language explanations (NLEs) and counterfactual examples (CFEs).

Key Contributions

1. First systematic study of quantization's effect on self-explanations, covering both explanation quality and explanation faithfulness across five datasets (eSNLI, HealthFC, ECQA, AG News, IMDb) and six models from the Llama3 (8B, 70B) and Qwen2.5 (7B, 14B, 32B, 72B) families.

2. A multi-perspective evaluation framework combining automatic metrics (BARTScore, TIGERScore for NLEs; label flip rate, perplexity, and textual similarity for CFEs), two NLE faithfulness metrics (biasing features and CC-SHAP), self-consistency checks for CFEs, an LLM-as-a-Judge evaluation, and a human user study.

3. A user study with 48 participants rating the trustworthiness and coherence of explanations from full-precision versus quantized Qwen2.5 models of sizes {7B, 32B, 72B}, plus a direct comparison showing that LLM-as-a-Judge evaluation fails to capture the quantization effects humans perceive.

4. Practical guidance for practitioners, including the finding that no single quantization technique consistently excels across task performance, explanation quality, and faithfulness, and the recommendation to validate SE quality per use case.

Main Findings

• Moderate but real quality degradation: Quantization typically leads to moderate declines in SE quality (up to 4.4%) and faithfulness (up to 3.9%). NLEs show greater sensitivity to quantization, while CFEs are relatively more robust.

• Human-perceived degradation is larger: The user study shows quantization considerably diminishes both the coherence and trustworthiness of SEs, by up to 8.5%. Full-precision models were perceived as producing more trustworthy and coherent explanations than their quantized counterparts.

• Counterfactual fluency suffers most: Quantization negatively affects CFE generation across validity, fluency, and textual similarity, with fluency most affected (on average 6.3%). Counterfactual faithfulness decreases by an average of 1.5% under quantization, and smaller LLMs show more noticeable faithfulness drops.

• NLE faithfulness declines modestly on average: Across faithfulness metrics, biasing features declined by 3.4% and CC-SHAP by 0.05. Faithfulness was preserved in the majority of cases. Llama3 models were more susceptible to quantization-induced degradation of NLE faithfulness than Qwen2.5 models, especially at 4-bit precision (e.g., Llama3-8B with GPTQ4 on HealthFC assessed by CC-SHAP).

• Larger models are not uniformly better at explanation quality: Larger quantized models do not consistently generate higher-quality SEs than smaller full-precision models, and lower-bit-precision LLMs do not consistently lag behind higher-bit-precision counterparts. The paper notes this contradicts a finding in Badshah and Sajjad (2024).

• Larger models preserve faithfulness better: NLEs generated by larger models tend to be more faithful, and larger models show greater robustness to quantization in preserving NLE faithfulness. SEs from larger quantized models are more frequently faithful than those from smaller full-precision models.

• No single quantization winner: No quantization method consistently excels across task performance, explanation quality, and faithfulness simultaneously. GPTQ8, AWQ, and bib8 were found to excel at preserving task performance, explanation quality, and faithfulness respectively. A trade-off exists between preserving SE characteristics and preserving task performance.

• Quantization can occasionally help: Quantization sometimes even improved self-explanation quality, which the authors attribute to reduced output-distribution entropy, more consistent and simple language use, or regularization and noise effects that limit weight precision.

• LLM-as-a-Judge misses the effect: Judge models generally agreed with each other (κ overall above 0.69, especially on coherence), but their agreement with human raters was consistently weak or occasionally negative and not statistically significant (all p-values > 0.05). This misalignment was particularly pronounced for CFEs. The paper suggests one explanation is that the full-precision versus quantized gap for CFEs is only 1.3%, so judges may converge on surface features such as fluency and length.

• Task performance and explanation quality are not tightly linked: Especially for larger models, quantization-induced task performance degradation generally did not contribute to NLE quality degradation, as indicated by weak or even negative correlations.

• Inter-annotator agreement: Krippendorff's α of 0.71 for NLEs and 0.64 for CFEs.

Methodology in Plain English

The researchers took six open-source LLMs and produced quantized versions of each using three common post-training quantization techniques: GPTQ (which uses second-order, Hessian-based optimization), AWQ (which handles activation outliers), and integer quantization implemented via BitsAndBytes. Quantized variants were at 4-bit and 8-bit widths (labeled bib4, bib8, gptq4, gptq8, and awq). All generations used temperature=0 for reproducibility, and automatic results were averaged over three runs with different seeds.

For each model variant, they generated two types of explanations. Natural language explanations were produced using ZeroCoT, a zero-shot prompting method that elicits step-by-step reasoning by adding "Let's think step by step." Counterfactual examples were produced using FIZLE, a two-stage process that first extracts salient keywords from the input and then uses them to guide counterfactual generation. Zero-shot settings were chosen deliberately to minimize confounding effects from quantization's known impact on in-context learning.

Explanation quality was scored automatically: NLEs with BARTScore (reference-based, comparing generated explanations to human-annotated ones) and TIGERScore (reference-free, using a fine-tuned Llama2 model to penalize errors in coherence, informativeness, and accuracy). CFEs were scored on label flip rate, perplexity (fluency), and textual similarity via token-level Levenshtein distance. Faithfulness was measured with biasing features and CC-SHAP for NLEs, and self-consistency checks for CFEs.

Finally, 48 native English speakers without explainability expertise participated in a user study. They rated trustworthiness and coherence on 5-point Likert scales. Sampling focused on eSNLI (the dataset overlapping NLEs and CFEs) and Qwen2.5 models of sizes {7B, 32B, 72B}; Qwen2.5-14B was excluded due to consistently suboptimal performance even without quantization. Thirty dataset indices were randomly sampled, each model-precision pair was evaluated by at least two annotators, and each annotator was assigned 15 explanations. The same data subset was used for an LLM-as-a-Judge comparison.

Why This Matters

Impact on research. The paper opens a new dimension in quantization research, which had previously examined multilingual capability, bias, alignment, and in-context learning, but not self-explanation. It also delivers a methodological warning: LLM-as-a-Judge evaluation did not reliably track human judgments of quantization's effect on explanations, and the paper recommends human evaluation where subjective judgment matters. It further shows that quantization affects quality and faithfulness differently, hypothesizing that under quantization, weights may be more sensitive than the decision boundary due to overparametrization.

Real-world applications:

• On-device and edge LLM deployment, where quantized models are standard and users may rely on the model's own explanations.
• High-stakes decision support, such as fact-checking (HealthFC) and medical or commonsense question answering (ECQA), where explanations drive user trust.
• Transparency and compliance reporting, where an explanation's perceived trustworthiness and coherence may matter as much as raw task accuracy.
• News classification and sentiment analysis pipelines (AG News, IMDb), where counterfactual explanations are used to probe model behavior.

Industry relevance. The paper's core practical message is that quantization remains a viable compression strategy, but its effect on explanations is highly context-dependent and can be sizable in specific configurations, so teams should empirically test multiple quantization strategies for their specific use case rather than assuming a one-size-fits-all answer. When faithfulness is the priority, the authors suggest considering larger quantized models rather than smaller full-precision ones.

Future Directions

• Extending beyond 4-bit and 8-bit. The authors explicitly list lower-bit quantization such as 1-bit or 2-bit as outside their scope, leaving open whether more aggressive compression further damages self-explanations.

• Investigating why specific configurations collapse. The paper notes that in some configurations the drop could be substantial and "not easily predictable" (for example, Qwen2.5-14B with gptq4 on eSNLI shows a biasing-feature faithfulness rate of 41.80 versus 90.00 at full precision), so identifying predictors of such failures is an open question.

• Disentangling quantization's effects from other model capabilities. The authors state that disentangling the impact of quantization from confounding factors is infeasible due to the black-box nature of LLMs, and they adopted established protocols instead.

• Improving automated evaluation of explanations. Given that LLM-as-a-Judge ratings diverged from human ratings and the authors tied this to the narrow 1.3% full-precision-versus-quantized gap for CFEs, better automated methods for detecting subtle explanation quality shifts are needed.

• Generalizing to other explanation types. The authors report an additional evaluation of feature attribution in the appendix to test generalizability beyond free-text self-explanations, suggesting further coverage of other SE types.

Target Audience

This paper is most valuable to researchers and practitioners working on explainable AI for LLMs, model compression and efficient inference, and evaluation methodology for generated text. It is also relevant to ML engineers deploying quantized models in settings where the model's own explanations are shown to users, and to teams that need to decide between model size and quantization level while preserving transparency. Readers should have an intermediate background in NLP model deployment and explanation evaluation to get the most from the results tables and faithfulness metrics.

Authors’ abstract

Quantization is widely used to accelerate inference and streamline the deployment of large language models (LLMs), yet its effects on self-explanations (SEs) remain unexplored. SEs, generated by LLMs to justify their own outputs, require reasoning about the model's own decision-making process, a capability that may exhibit particular sensitivity to quantization. As SEs are increasingly relied upon for transparency in high-stakes applications, understanding whether and to what extent quantization degrades SE quality and faithfulness is critical. To address this gap, we examine two types of SEs: natural language explanations (NLEs) and counterfactual examples, generated by LLMs quantized using three common techniques at distinct bit widths. Our findings indicate that quantization typically leads to moderate declines in both SE quality (up to 4.4%) and faithfulness (up to 3.9%). The user study further demonstrates that quantization considerably diminishes both the coherence and trustworthiness of SEs (by up to 8.5%). Compared to smaller models, larger models show limited resilience to quantization in terms of SE quality but maintain more faithfulness. Moreover, no quantization technique consistently excels across task accuracy, SE quality, and faithfulness. Because quantization's impact varies considerably by context and can be sizable in specific cases, we recommend validating SE quality for the intended use case. Despite these sometimes considerable drops, quantization remains an effective compression technique when its impact on SEs is properly validated.