Evaluation of Embedding-Based and Generative Methods for LLM-Driven Document Classification: Opportunities and Challenges

Rong Lu; Hao Liu; Song Hou

cs.IRApr 5, 2026

Evaluation of Embedding-Based and Generative Methods for LLM-Driven Document Classification: Opportunities and Challenges

Authors: Rong Lu, Hao Liu, Song Hou

Abstract

This work presents a comparative analysis of embedding-based and generative models for classifying geoscience technical documents. Using a multi-disciplinary benchmark dataset, we evaluated the trade-offs between model accuracy, stability, and computational cost. We find that generative Vision-Language Models (VLMs) like Qwen2.5-VL, enhanced with Chain-of-Thought (CoT) prompting, achieve superior zero-shot accuracy (82%) compared to state-of-the-art multimodal embedding models like QQMM (63%). We also demonstrate that while supervised fine-tuning (SFT) can improve VLM performance, it is sensitive to training data imbalance.

Figures & tables

Model	Intra ( $\downarrow$ )	Inter ( $\uparrow$ )	Ratio ( $\uparrow$ )	Silh. ( $\uparrow$ )	DB ( $\downarrow$ )	CH ( $\uparrow$ )	F1 ( $\uparrow$ )	Acc. ( $\uparrow$ )
QQMM-embed	0.088	0.161	1.822	0.210	2.180	239.537	0.64	0.63
gme-Qwen2-VL-7b	0.128	0.098	0.761	0.074	3.361	95.294	0.59	0.62
mmE5-mllama-11b	0.143	0.112	0.785	0.089	3.286	95.304	0.51	0.53
vdr-2b-multi-v1	0.208	0.167	0.804	0.068	4.107	89.497	0.37	0.38
clip-ViT-L-14	0.205	0.110	0.536	0.002	5.231	65.732	0.18	0.22

Table 1: Performance of five multimodal embedding models on the benchmark dataset. Arrows (

\downarrow,\uparrow

) indicate whether lower or higher values are preferable. QQMM performs better than GME [ 10 ] , mmE5 [ 3 ] , vdr [ 5 ] , and CLIP [ 7 ] .

Model	Base Prompt		Plus Prompt
Model	Accuracy	F1	Accuracy	F1
Qwen2.5-VL-72B	0.78	0.77	0.82	0.82
Qwen2.5-VL-7B	0.65	0.65	0.76	0.75
Gemma-3-27B	0.64	0.65	0.70	0.69
Mistral-3.2-24B	0.55	0.55	0.58	0.55

Table 2: Zero-shot classification performance of various VLMs. The “plus” prompt includes CoT and domain definitions while the “base” version mainly utilizes persona prompting.

Figure 1: Classification performance for both embedding and VLMs under various configurations. VLMs generally outperform embedding models, especially after fine-tuned with sufficient training samples. Dots are individual models’ results.

Earth Science

See all 23 papers

The Embedder's Dilemma: LLMs Are Better, but at What Cost?
Aug 13, 2026Adnan El Assadi, Niklas Muennighoff, Jinhyuk LeeLLM Inference Efficiency LLM Evaluation
Harvard University · Stanford University · Independent Researcher
Beyond Accuracy: Robustness, Cost, and Governance Trade-offs for Vision-Language Models in Templated Document Extraction
Sep 14, 2026Kushal Patel, Pushkal Shrivastava, Mackenzie Lees +4VLM Evaluation VLM Robustness
John Hancock, 200 Berkeley St, Boston, MA 02116, USA
Image Recognition with Vision and Language Embeddings of VLMs
Sep 11, 2025Illia Volkov, Nikita Kisel, Klara Janouskova +1Image Classification VLM Evaluation
Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague Prague, Czechia

Aug 13, 2026J/K move · Enter open · S save

The Embedder's Dilemma: LLMs Are Better, but at What Cost?

Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee

Harvard University · Stanford University · Independent Researcher

LLM Inference Efficiency LLM Evaluation Text Embedding Evaluation Information Retrieval Text Embedding Models

Open paper AskPDF arXiv

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.

Abstract

Earth Science

Explore similar work

The Embedder's Dilemma: LLMs Are Better, but at What Cost?

Beyond Accuracy: Robustness, Cost, and Governance Trade-offs for Vision-Language Models in Templated Document Extraction

Image Recognition with Vision and Language Embeddings of VLMs