Large language models offer far more stable explanations than traditional text classifiers, study finds

As organizations increasingly deploy NLP systems in regulated sectors, they need explainable AI that users can trust—but validating whether those explanations remain reliable under real-world conditions has proven challenging, especially when systems are accessed only through APIs.

A new study from researchers including Guilin Zhang tackles this problem by developing a framework to measure how robust explanations are when inputs get scrambled, deleted, or otherwise disrupted in ways users might encounter.

The robustness test#

The team designed a black-box evaluation protocol based on leaving out individual tokens one at a time to identify which words matter most to a model's decision. They then introduced realistic disturbances—word swaps, deletions, shuffling, and back-translation—at varying intensity levels across three benchmark datasets, testing six models from both the encoder family (BERT, RoBERTa) and the newer decoder approach (Qwen, Llama).

The scale of the comparison was substantial: the analysis covered over 64,000 test cases. Robustness was measured by tracking how often the model's top-ranked explanation tokens changed when inputs were perturbed.

Decoder models win decisively#

The results strongly favor large language models built on a decoder architecture. These models produced explanations that flipped far less frequently—73% lower flip rates on average compared to the older encoder classifiers. The advantage also grew with model size: scaling from 7 billion to 70 billion parameters yielded a 44% improvement in stability.

Practical tradeoffs for deployment#

Beyond raw performance, the research identified the cost-benefit relationship between inference expense and explanation robustness. That tradeoff analysis gives organizations a practical guide for choosing which models and explanation methods fit their needs before pushing systems into production, particularly in compliance-heavy industries where explanation quality matters legally and operationally.