A Table Is Worth 64 Tokens: Pixel-level Compression for Multi-Table Document Question Answering

Iñigo Alonso; Mirella Lapata

cs.AIAug 27, 2026

A Table Is Worth 64 Tokens: Pixel-level Compression for Multi-Table Document Question Answering

Authors: Iñigo Alonso, Mirella Lapata

Organizations: School of Informatics University of Edinburgh Edinburgh, UK

Abstract

Answering questions over real-world documents requires processing long inputs that interleave text with tables. Optical context compression, which represents context as images, promises to reduce token cost, but its effect on table understanding remains unclear. We study pixel-level table compression for question answering over documents with multiple tables, evaluating five VLMs across two benchmarks and five visual-token budgets. Representing tables as images at native resolution matches text in both performance and efficiency, but downscaling them makes models compensate the loss in readability with longer, less effective reasoning traces that cancel the expected savings. Highly downscaled tables, however, preserve enough signal to identify whether they are relevant to a question. We exploit this asymmetry with a training-free, two-step method: the model first identifies the tables needed to answer a question from a pixel-compressed context, and then reasons over those at native resolution. On long documents, our method saves 41% of total tokens and gains 7 accuracy points over single-step QA with native resolution tables. It also uses 15% fewer tokens than the most efficient single-step compressed configuration, with no accuracy loss.

Figures & tables

Figure 1: Our two-stage identify first, reason later method. Stage 1 : the model receives the document with every table compressed to a small visual token budget and, instead of answering, identifies the tables it needs. Stage 2 : the identified tables are supplied at native resolution and the model reasons over them to answer.

	MultiHiertt	FinLongDocQA
Examples	1,044	400
Tables / doc	3 [2, 7]	80 [3, 122]
Pages / doc	4 [2, 8]	100 [16, 156]
Max. Tokens (HTML)	12,264	127,816
Evidence tables	1 [0, 3]	2 [1, 8]
Evidence pages	1 [1, 4]	2 [2, 3]

Table 1: Statistics of the evaluation subsets (mode, with [min, max] intervals, except ‘Max. Tokens’, which reports the true maximum). FinLongDocQA annotates evidence at the page level; we treat every table on an evidence page as an evidence table. Answer in ‘Text/Table/Text and Table’ gives the percentage of questions answerable from text alone, from a table alone, or both.

Model	Tokenization	Context	Reasoning
Gemma 4 E4B	Fixed	128k	✓
Gemma 4 26B-A4B	Fixed	256k	✓
Qwen 3 VL-8B Instruct	Variable	256k	✗
Qwen 3 VL-8B Thinking	Variable	256k	✓
Qwen 3.5-9B	Variable	256k	✓

Table 2: The five open-weight VLMS we evaluate. Tokenisation is whether the visual encoder emits a fixed or variable number of visual tokens per image; Context is the supported context-window length; Reasoning ( ✓ / ✗ ) marks test-time reasoning mode.

Figure 2: Table transcription performance against input tokens on MultiHiertt. The html

\rightarrow

html control is nearly perfect, while latex

\rightarrow

html shows that format conversion introduces error even without visual perception. Transcription from images degrades progressively as the visual-token budget decreases.

Panel of Figure 3 — Figure 3: QA accuracy on MultiHiertt (test-time reasoning enabled) for Gemma 4 26B-A4B and Qwen 3.5-9B, plotted against median input tokens (left) and generated tokens (right). At large budgets, table images match HTML accuracy (left) using fewer input tokens; stronger compression reduces accuracy and produces longer reasoning traces (right). Results for all models and FinLongDocQA are in Appendix E .

Figure 4: Relevant-table identification (F1) for MultiHiertt across token budgets and table representations (Gemma 4-26B, Qwen 3.5-9B). Bars (y axis), decompose F1 into achieved vs. shortfall on requested tables; html and tableid are references; the horizontal lines mark random choice and ColQwen2@1, its best-F1 operating point. The percentage above each bar is the overall downstream QA accuracy, so identification gains carry through to the downstream task, not just retrieval (see Appendix E for more detail).

Panel of Figure 5 — Figure 5: Accuracy against total tokens for direct QA and our two-step method, on MultiHiertt (top) and FinLongDocQA (bottom), for Gemma 4 26B-A4B and Qwen 3.5-9B. The two left columns feed the full document; the two right columns feed a context retrieved by BGE-M3, with single-step QA over the retrieved units at increasing $k$ as reference. Markers along each line correspond to visual token budgets, and bands show $\pm 1$ standard error of the mean across examples (Appendix B ). Results for all models are provide in Appendix E .

MultiHiertt
Identify first (64)	Matched	Image		Retrieval
Reason later vs.	Tok.	Tok.	Acc.	Tok.	Acc.
Qwen 3.5-9B	14.0	20.9	0.0	25.2	$-$ 0.7
Qwen 3 VL-8B	7.8	14.8	$-$ 7.8	17.3	$-$ 2.7
Gemma 4 26B	8.8	43.8	$-$ 8.0	35.9	$-$ 8.0
Gemma 4 E4B	$-$ 6.2	35.2	$-$ 1.8	24.1	$-$ 3.4

Table 3: Percentage of tokens saved (Tok.) and accuracy difference in points (Acc.; positive favours our method) of our two-stage method against three single-step references. Matched is the cheapest single-step configuration our method matches in accuracy, judged by the confidence intervals of Appendix C ; Image is single-step QA over table images at native resolution; Retrieval is single-step QA over a context retrieved by BGE-M3.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Prompt used for table transcription.

Figure 8: Prompt used for our two-stage method.

Panel of Figure 9 — Figure 9: Direct QA with and without test-time reasoning across visual token budgets on MultiHiertt, for all four reasoning-capable models. Compression widens the gap between the two conditions, as lower budgets induce longer reasoning traces.

Panel of Figure 10 — Figure 10: Accuracy against total tokens for direct QA and our two-step method, on MultiHiertt (top) and FinLongDocQA (bottom), for Qwen 3 VL 8B, Qwen 3.5 9B, Gemma 4 E4B, and Gemma 4 26B-A4B for full document. Markers along each line correspond to visual token budgets, and bands show $\pm 1$ s.e.m. across examples (Appendix B ).

Model	Budget	Task	Acc.	Input text	Input visual	Output text	Total
Qwen 3 VL-8B	-	Dir. QA	24.8	4696	0	32768	36752
Qwen 3 VL-8B	1024	Dir. QA	42.2	2601	1351	2745	6990
Qwen 3 VL-8B	512	Dir. QA	40.0	2601	1089	2563	6682
Qwen 3 VL-8B	256	Dir. QA	34.5	2601	722	3100	6574
Qwen 3 VL-8B	128	Dir. QA	28.2	2601	430	3526	7053
Qwen 3 VL-8B	64	Dir. QA	17.8	2601	294	4522	7346

Appendix

Table 4: Full-context results on MultiHiertt. Budget is the visual token budget applied to each table image, with ‘–’ denoting the html setting in which tables are serialized as text; Task is the QA setting. Acc. is exact-match accuracy; the remaining columns report median token counts per example, split into input text tokens, input visual tokens, and generated tokens, with their sum in Total .

Model	Budget	Task	Acc.	Input text	Input visual	Output text	Total
Qwen 3.5-9B	-	Dir. QA	29.0	102782	0	4760	108016
Qwen 3.5-9B	1024	Dir. QA	25.8	61050	33444	3690	100936
Qwen 3.5-9B	512	Dir. QA	25.5	61050	24916	3844	91947
Qwen 3.5-9B	256	Dir. QA	25.0	61050	15950	4116	82064
Qwen 3.5-9B	128	Dir. QA	19.0	61050	8914	5006	75332
Qwen 3.5-9B	64	Dir. QA	14.8	61050	6064	5252	73452

Appendix

Table 5: Full-context results on FinLongDocQA. Budget is the visual token budget applied to each table image, with ‘–’ denoting the html setting in which tables are serialized as text; Task is the QA setting. Acc. is exact-match accuracy; the remaining columns report median token counts per example, split into input text tokens, input visual tokens, and generated tokens, with their sum in Total . Gemma 4 E4B and Qwen3-VL-8B are omitted, as this dataset exceeds the context window of the former and makes every configuration considerably more expensive to run.

Model	Budget	Task	Acc.	Input text	Input visual	Output text	Total
Qwen 3 VL-8B	1024@100	Re. QA	39.0	2583	1358	2828	7005
Qwen 3 VL-8B	1024@50	Re. QA	39.4	1927	1348	2901	6372
Qwen 3 VL-8B	1024@25	Re. QA	39.5	1091	1260	2679	5153
Qwen 3 VL-8B	1024@5	Re. QA	37.5	358	768	2934	4066
Qwen 3 VL-8B	1024@3	Re. QA	36.9	290	616	2952	3936
Qwen 3 VL-8B	1024@2	Re. QA	32.8	255	531	3188	3940

Appendix

Table 6: Results on MultiHiertt with a context retrieved by BGE-M3. Budget reports the visual token budget applied to each table image and the number of retrieved units

k

, in the form budget@

k

; Task is the QA setting. Acc. is exact-match accuracy; the remaining columns report median token counts per example, split into input text tokens, input visual tokens, and generated tokens, with their sum in Total .

Model	Budget	Task	Acc.	Input text	Input visual	Output text	Total
Qwen 3.5-9B	1024@200	Re. QA	33.2	14574	24480	3188	44049
Qwen 3.5-9B	1024@100	Re. QA	29.5	6266	16518	3908	28185
Qwen 3.5-9B	1024@10	Re. QA	24.2	782	1870	6042	9134
Gemma 4-26B A4B	1024@200	Re. QA	42.8	14688	64844	5413	88282
Gemma 4-26B A4B	1024@100	Re. QA	41.2	6370	42202	6561	57069
Gemma 4-26B A4B	1024@10	Re. QA	30.8	804	4351	8650	14367

Appendix

Table 7: Results on FinLongDocQA with a context retrieved by BGE-M3. Budget reports the visual token budget applied to each table image and the number of retrieved units

k

, in the form budget@

k

Figure 12: Relevant-table identification (F1) on MultiHiertt across token budgets and table representations, for all models. Bars show identification F1 (y axis), decomposed into F1 achieved vs. shortfall on requested tables; HTML and tableid are references, and the horizontal lines mark random choice and ColQwen2@1, its best-F1 operating point. The percentage above each bar is the overall downstream QA accuracy.

Figure 13: Relevant-table identification (F1) on FinLongDocQA for the two models evaluated on this dataset. Bars and references are as in Figure 12 , except that the retrieval line marks ColQwen2@4, its best-F1 operating point on this dataset. FinLongDocQA annotates evidence at the page level, so every table on an evidence page counts as relevant and the resulting scores are an upper bound on true identification performance. These results are shown for reference and are not used to support any claim in the paper.

Figure 14: Accuracy against input tokens (left) and generated tokens (right) as a function of the budget at which the requested table is returned, for Qwen 3.5-9B on MultiHiertt. Colours distinguish the visual token budget applied to the context, and each labelled point is the budget at which the requested table is returned ( html , or images at 1,024 to 64 tokens).

Figure 15: Retrieval performance on MultiHiertt (left) and FinLongDocQA (right), reporting average recall (top row) and average F1 (bottom row) against the number of retrieved units. Text is BGE-M3 and image is ColQwen2. In any (paragraph) and any (table) the pool contains both paragraphs and tables, and only the text retriever applies; in tables the pool contains tables alone and both retrievers are compared. FinLongDocQA omits any (paragraph) , as its evidence is annotated at the page level.

Panel of Figure 16 — Figure 16: Accuracy against total tokens for our two-stage method with model-based table identification, and for direct QA over a context in which the tables are supplied by ColQwen2 at increasing $k$ , on MultiHiertt (top) and FinLongDocQA (bottom). Both conditions receive the full document text; they differ only in how the tables reach the model. Markers along our curves correspond to visual token budgets, and bands show $\pm 1$ s.e.m. across examples (Appendix B ).

Document QA

Document question answering uses AI to answer questions from text, tables, images, and structured content in documents.

See all 92 papers

DocPrune:Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning
Apr 24, 2026Joonmyung Choi, Sanghyeok Lee, Jongha Kim +4Efficient VLM Inference Document QA
Korea University · KAIST · Adobe Research
RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation
Aug 1, 2026Jiayang Yu, Jialun Zhong, Lei ZouMultimodal RAG Retrieval-Augmented Generation
Wangxuan Institute of Computer Technology, Peking University Beijing, China
Tables Decoded: DELTA for Structure, TARQA for Understanding
Sep 15, 2026Jahanvi Rajput, Dhruv Kudale, Saikiran Kasturi +2Table QA Table Structure Recognition
Indian Institute of Technology Bombay · IIT Bombay · BharatGen

Apr 24, 2026J/K move · Enter open · S save

DocPrune:Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning

Joonmyung Choi, Sanghyeok Lee, Jongha Kim, Sehyung Kim+3

Korea University · KAIST · Adobe Research

Efficient VLM Inference Document QA Multimodal Token Compression Visual Token Pruning Document Image Understanding

Open paper AskPDF arXiv

Recent advances in vision-language models have demonstrated remarkable performance across diverse multi-modal tasks, including document question answering that leverages structured visual cues from text, tables, and figures. However, unlike natural images, document images contain large backgrounds and only sparse supporting evidence, leading to the inefficient consumption of substantial computational resources, especially for long documents. We observe that existing token-reduction methods for natural images and videos fall short in utilizing the structural sparsity unique to documents. To address this, we propose DocPrune, a training-free and progressive document token pruning framework designed for efficient long-document understanding. The proposed method preserves only the essential tokens for the task while removing unnecessary ones, such as background or question-irrelevant tokens. Moreover, it automatically selects the appropriate layers to initiate token pruning based on the model's level of comprehension. Our experiments on the M3DocRAG show that DocPrune improves throughput by 3.0x and 3.3x in the encoder and decoder, respectively, while boosting the F1 score by +1.0, achieving both higher accuracy and efficiency without any additional training.

Abstract

Document QA

Explore similar work

DocPrune:Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning

RAGOCR: Optical Compression of Retrieval-Augmented Text via Visual Representation

Tables Decoded: DELTA for Structure, TARQA for Understanding