Can Large Language Models Still Explain Themselves? Investigating the Impact of Quantization on Self-Explanations
This paper investigates how quantization affects large language models' self-explanations, finding that while it generally causes moderate declines in explanation quality, faithfulness, and user trust, the impact varies by context and model size, necessitating case-specific validation before deployment.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Large language models are powerful tools that can write, reason, and explain their own thinking. When these systems make a decision, they can often generate a sentence or two describing why they chose that path, a feature known as a self-explanation. This ability is becoming increasingly important for high-stakes situations, such as medical diagnosis or legal analysis, where users need to trust that the machine is reasoning correctly and not just guessing. To make these massive systems faster and cheaper to run on everyday devices, engineers often use a technique called quantization. This process is like compressing a high-resolution photograph into a smaller file size; it reduces the precision of the numbers inside the model to save space and speed up calculations, but it risks losing some of the fine detail that makes the image clear. The critical question for researchers has been whether this compression also blurs the model's ability to explain itself truthfully.
A team of scientists set out to investigate exactly this problem. They took several large language models of different sizes and applied three common compression techniques to them, creating versions that used fewer bits of information to store their knowledge. They then asked these models to generate two types of explanations for various tasks: natural language justifications, which are written sentences explaining a decision, and counterfactual examples, which are slightly modified versions of an input that would cause the model to change its mind. By comparing the compressed models against their original, uncompressed versions, the researchers could see if the act of shrinking the model also shrank its capacity for honest reasoning.
The results showed that while compression does have an effect, it is not as catastrophic as one might fear, though it is not without cost. The study found that quantization typically leads to a moderate decline in the quality of the explanations, with scores dropping by up to 4.4 percent, and a similar decrease in faithfulness, meaning the explanations became slightly less aligned with the model's actual internal logic, dropping by up to 3.9 percent. However, the human experience of these changes was more pronounced. In a study involving forty-eight people who read and rated the explanations, the compressed models were perceived as significantly less trustworthy and less coherent, with trustworthiness scores falling by as much as 8.5 percent. This suggests that while the models might still be technically functional, the explanations they produce feel less reliable to a human reader.
Interestingly, the researchers discovered that size does not always guarantee resilience. While larger models were generally better at maintaining the truthfulness of their explanations even when compressed, they did not necessarily produce higher-quality text than smaller, uncompressed models. Furthermore, no single compression method proved to be the best across the board; some techniques preserved task accuracy better, while others were slightly better at keeping explanations coherent. The study also highlighted a limitation in how we currently evaluate these systems. The researchers found that using another artificial intelligence to judge the quality of these explanations failed to match human judgment. The automated judges often missed the subtle ways in which compression degraded the trustworthiness of the text, a gap that only human evaluation could reveal.
Ultimately, the paper concludes that quantization remains a viable strategy for deploying these models, provided that developers test the specific explanations their models will generate before putting them to work. The degradation in performance is often small enough to be acceptable, but in situations where transparency is critical, such as healthcare or justice, the slight loss in coherence and trustworthiness matters. The findings suggest that there is no one-size-fits-all solution; instead, practitioners must carefully validate how compression affects the specific type of explanation their system will provide, ensuring that the model remains a trustworthy partner in decision-making even after it has been compressed.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.