When Does Smaller Beat Larger? Accuracy Equivalence and Output Modality Advantage of Encoder-Only Models in Domain Question Answering

Abstract

In domain-specific knowledge question answering (QA) for fields such as oceanography, systems face not only deployment constraints--offline operation, data sovereignty, and long-term cost control--but also a deeper, often overlooked problem: generative large language models (LLMs), as pure text generators, are architecturally incapable of outputting rich media content (images, videos, document links) verified by domain experts, yet such content is a core requirement of domain QA. Taking ocean-domain knowledge services as the background, this paper systematically investigates the applicability conditions of lightweight encoder-only models for domain QA from five dimensions--accuracy, latency, cost, offline capability, and model size--combined with the structural constraint of output modality. Experiments show that in closed-domain scenarios with fixed, well-bounded answers, a BERT model with only approximately 102 million parameters, combined with high-quality training data, achieves 96.79% accuracy on the validation set. This is practically equivalent to traditional methods (SVM+TF-IDF, TextCNN) as confirmed by McNemar's test (p > 0.05), and also practically equivalent to the zero-shot classification accuracy of the hundred-billion-parameter DeepSeek (94.80%). Meanwhile, BERT comprehensively outperforms LLMs in inference latency (approximately 100 ms vs. 2-5 s), cost, and offline deployability. More critically, this paper formally reveals a structural blind spot of LLMs: their output endpoints cannot, by architecture, implement verified rich media embedding, whereas encoder-only models with preset answer banks natively support this capability. The core contributions include: (1) a confidence-driven semantic deduplication method for multi-source knowledge; (2) a systematic edge-side inference performance benchmark; (3) a same-data comparison between traditional methods and fine-tuned BERT with McNemar statistical tests; (4) the output modality advantage of encoder-only models and a formal analysis of the structural blind spot in generative architectures; (5) a domain QA model selection decision framework integrating five quantitative evaluation dimensions with output modality qualitative constraints.