Paper deep dive

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

Iñigo Alonso, Mirella Lapata

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 80

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 8/28/2026, 4:26:45 AM

Summary

This paper investigates pixel-level compression for multi-table document question answering (QA). The authors evaluate five Vision-Language Models (VLMs) across two benchmarks (MultiHiertt and FinLongDocQA) and five visual-token budgets. They find that while native resolution table images match text in performance and efficiency, aggressive downscaling degrades fine-grained reading but preserves enough coarse signal for table identification. Exploiting this asymmetry, they propose a training-free, two-step method: first identifying relevant tables from a compressed context, then reasoning over those specific tables at native resolution. This method saves 41% of total tokens and improves accuracy by 7 points over single-step QA with native resolution tables on long documents, while using 15% fewer tokens than the most efficient single-step compressed configuration with no accuracy loss.

Entities (10)

FinLongDocQA · dataset · 99%MultiHiertt · dataset · 99%Pixel-level Compression · method · 98%Two-Stage Method · method · 97%Gemma 4 · model · 95%Qwen 3.5 · model · 95%Qwen-3-VL · model · 95%Accuracy Gain · metric · 92%Token Savings · metric · 92%Visual Token Budget · hyperparameter · 90%

Relation Signals (8)

MultiHiertt → usedfor → Short-context setting

confidence 95% · We use MultiHiertt as a short-context setting, with typically three tables and four pages per document

FinLongDocQA → usedfor → Long-context setting

confidence 95% · and FinLongDocQA as a long-context setting, with typically 80 tables and 100 pages per document.

Two-Stage Method → uses → Pixel-level Compression

confidence 95% · 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

Two-Stage Method → outperforms → Single-step QA

confidence 93% · On long documents, our method saves 41% of total tokens and gains 7 accuracy points over single-step QA with native resolution tables.

Pixel-level Compression → degrades → Fine-grained reading

confidence 90% · Lower resolutions degrade fine-grained reading and induce longer reasoning traces

Gemma 4 → evaluatedon → MultiHiertt

confidence 90% · Figure 2: Table transcription performance against input tokens on MultiHiertt... Gemma 4-26B

Qwen-3.5 → evaluatedon → MultiHiertt

confidence 90% · Figure 2: Table transcription performance against input tokens on MultiHiertt... Qwen 3.5-9B

Pixel-level Compression → →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

76,256 characters extracted from source content.

Expand or collapse full text

A Table Is Worth 64 Tokens: Pixel-level Compression for Multi-Table Document Question Answering Iñigo Alonso and Mirella Lapata School of Informatics University of Edinburgh Edinburgh, UK ialonso, mlap@ed.ac.uk Abstract Answering questions over real-world docu- ments requires processing long inputs that in- terleave text with tables. Optical context com- pression, 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 an- swering 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 effec- tive 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 asym- metry with a training-free, two-step method: the model first identifies the tables needed to answer a question from a pixel-compressed con- text, and then reasons over those at native reso- lution. On long documents, our method saves 41% of total tokens and gains 7 accuracy points over single-step QA with native resolution ta- bles. It also uses 15% fewer tokens than the most efficient single-step compressed configu- ration, with no accuracy loss. 1 1 Introduction Real-world documents are long and heterogeneous, mixing free-form text with structured elements such as tables. Despite advances in efficient at- tention, processing long inputs remains computa- tionally costly (Liu et al., 2025), and performance degrades once contexts approach or exceed the lengths seen during pre-training (Liu et al., 2024; Hsieh et al., 2024). Representing the input effi- ciently is therefore not just an optimization goal but a requirement for handling documents at scale. 1 Code and data will be released upon publication. Recent work on optical context compression has shown that text rendered as an image typically con- sumes fewer tokens than the same text represented as raw text (Wei et al., 2025). Originally tested on OCR-style reconstruction, these findings have since been extended to downstream tasks, both with dedicated training (Xing et al., 2025a; Cheng et al., 2026) and with off-the-shelf VLMs (Li et al., 2025). This evidence, however, is limited to plain text which motivates our central question: can pixel- level table compression improve efficiency at no performance cost? To answer it, we focus on question answering (QA) over financial documents, a task representa- tive of real-world challenges: long context, hetero- geneous information (text and tables), and ques- tions requiring numerical reasoning. We evaluate 5 open-weight models spanning fixed- and native- resolution visual encoders, under a common frame- work of five visual token budgets (64, 128, 256, 512, and 1,024 tokens per image), using the num- ber of tokens processed by the decoder backbone as a proxy for efficiency. Our experiments show that simply lowering the resolution is not enough: under aggressive compression, models cannot de- cipher fine-grained cell values, and compensate by reasoning more, to the point where the extra rea- soning eats into the input savings that motivated the compression in the first place. At the same budgets, however, models remain able to identify which tables are needed to answer a question. Drawing on these findings, we propose a simple two-stage method (see Figure 1). The model first receives the document with every table compressed to a given budget and, instead of answering directly, it identifies the relevant tables first; these tables are then supplied as images at native resolution, and the model reasons over them to produce an answer. Concurrent work shows that decoupling evidence identification from reasoning can improve long- context reasoning (Guan et al., 2026; Zhao et al., arXiv:2608.26949v1 [cs.AI] 27 Aug 2026 Stage 1: Identify evidence over compressed tables Stage 2: Reason over requested tables in full-resolution 64 tok 64 tok TABLE 1 TABLE 2 request:T2 Model Model "...margins in Table 2: A=28%, B=34%,C=31%" → B is highest..." answer:2009 full resolution TABLE 2 64 tok 64 tok TABLE 1 TABLE 2 + (no reasoning) (reasons) Q: In which year is Commercial Utilized Loans and leases positive? 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. 2026). Our method shows that compression makes the identification step cheap, since it preserves the coarse information needed to locate relevant tables. Our contributions can be summarised as follows: •We characterize the accuracy-efficiency trade-off of pixel-level table compression in multi-table document QA, across 5 VLMs, two datasets, and five visual-token budgets. At native resolution, table images match their HTML serialization in accuracy while using fewer tokens; under aggres- sive compression, however, the rendered tables become illegible, models can no longer read in- dividual cell values, and accuracy drops. • We identify an asymmetry in how compression affects model behaviour. Lower resolutions de- grade fine-grained reading and induce longer reasoning traces that offset part of the input- token savings, yet even heavily downscaled ta- bles retain enough coarse signal for question- conditioned identification of the relevant tables. • We exploit this asymmetry with a two-stage method that follows an identify first, reason later recipe: it first identifies relevant tables in a com- pressed context and then reasons over only those tables at native resolution. On long documents it saves 41% of total tokens while gaining 7 ac- curacy points over single-step QA with native- resolution tables, and matches the accuracy of the most efficient single-step compressed config- uration using 15% fewer tokens. •We situate the method against external retrieval. A dedicated retriever identifies evidence tables more accurately than a model reading a com- pressed context, but this advantage does not carry through to a better accuracy-efficiency trade-off; our method matches or exceeds retrieval at lower token cost, and continues to save tokens when applied on top of retrieved context. 2 Related Work Efficient Long-context InferenceLong-context models remain costly and often struggle to use ev- idence distributed across long inputs (Liu et al., 2025, 2024). Complementary approaches reduce the context processed in each forward pass through prompt compression (Yoon et al., 2024), memory (Chen et al., 2024c), agentic reading (Zhang et al., 2024), or retrieval (Günther et al., 2025). Recent work further shows that separating evidence iden- tification from reasoning improves long-context performance (Guan et al., 2026; Zhao et al., 2026). Our method adopts this separation but uses pixel compression to make evidence identification over multi-table documents inexpensive; it can be ap- plied to either full or retrieved contexts. Optical Context CompressionRecent work has explored representing language through rendered pixels (Rust et al., 2023; Lee et al., 2023; Lotz et al., 2023; Gao et al., 2024). A related line treats this representtion as a form of context compression, since visual encoders may represent rendered text with fewer tokens (Wei et al., 2025). This idea has been extended to downstream tasks through spe- cialised training (Wang et al., 2024a; Xing et al., 2025a; Cheng et al., 2026) and, more recently, with off-the-shelf VLMs (Li et al., 2025). Related ap- proaches prune visual tokens (Chen et al., 2024b; MultiHierttFinLongDocQA Examples1,044400 Tables / doc3 [2, 7]80 [3, 122] Pages / doc4 [2, 8]100 [16, 156] Max. Tokens (HTML)12,264127,816 Evidence tables1 [0, 3]2 [1, 8] Evidence pages1 [1, 4]2 [2, 3] Answer in Text11.0%0.0% Answer in Table32.4%100.0% Answer in Text and Table56.6%0.0% Questions / doc3 [2, 6]1 [1, 6] Table 1: Statistics of the evaluation subsets (mode, with [min, max] intervals, except ‘Max. Tokens’, which re- ports the true maximum). FinLongDocQA annotates evidence at the page level; we treat every table on an evidence page as an evidence table. Answer in ‘Tex- t/Table/Text and Table’ gives the percentage of questions answerable from text alone, from a table alone, or both. Shang et al., 2025; Choi et al., 2026), but are of- ten tied to particular architectures. In contrast, we study training-free pixel compression for struc- tured, information-dense tables across both fixed- and variable-resolution visual encoders. Table Representation and Reasoning Tables are commonly linearised for language models (Herzig et al., 2020; Sui et al., 2023; Wang et al., 2024b; Zou et al., 2025), although linearisation can impair structural reasoning over long tables (Wang et al., 2025). Other work compares visual and tex- tual table representations (Deng et al., 2024; Zheng et al., 2024; Alonso et al., 2024; Kim et al., 2024; Singh et al., 2025) or dynamically routes between them (Xing et al., 2025b; Wu et al., 2025). Most closely related, Kwok et al. (2026) dynamically select visual or textual representations for portions of a table during reasoning. Their method targets latency in single-table, multi-turn QA at a fixed image resolution; we instead study token efficiency under controlled compression in documents con- taining multiple interleaved tables and introduce a training-free method that separates table identifica- tion from full-resolution reasoning. 3 Experimental Framework 3.1 Datasets Existing table-understanding benchmarks (Chen et al., 2020; Parikh et al., 2020; Kim et al., 2024; Alonso et al., 2026) often present tables in isola- tion, whereas real-world document QA requires models to identify and reason over relevant tables embedded within surrounding text. Our study also requires tables that can be represented faithfully as both text and images, so that differences between modalities reflect the representation rather than dif- ferences in content. We therefore focus on bench- marks that contain multiple tables interleaved with text and provide tables in a structured form that can be rendered into images with content parity. These criteria lead us to two benchmarks: Mul- tiHiertt (Zhao et al., 2022) and FinLongDocQA (Wang et al., 2026). Both combine multiple tables with surrounding text, but differ substantially in scale. We use MultiHiertt as a short-context set- ting, with typically three tables and four pages per document, and FinLongDocQA as a long-context setting, with typically 80 tables and 100 pages per document. Dataset statistics are shown in Table 1. MultiHiertt test labels are private, so we evaluate on the development set. Although current models may have encountered this split during pre-training, our analysis compares configurations within each model rather than absolute performance across models. For FinLongDocQA, we evaluate 400 randomly sampled questions from the table cate- gory, ensuring that the answer evidence is grounded in tables. 2 We sample from documents whose HTML serialization fits within 128k tokens un- der the Qwen 3.5 tokenizer, allowing all configura- tions to be evaluated within the models’ supported context windows. We report confidence intervals throughout. 3.2 Input Settings We compare two input settings that differ only in how tables are represented; the document, question, and the prompt are always provided as text. Text (HTML) Tables are serialized as HTML, which is a strong textual representation that pre- serves hierarchical structure, including multi-level headers and spanning cells (Sui et al., 2023). This setting serves as the performance reference for both accuracy and token usage. Hybrid (Text+Images) Document text is pro- vided as text, while tables are rendered as images from the same HTML used in the textual setting. This gives the two modalities identical content and structure, allowing us to isolate the effect of repre- 2 At the time of dataset selection, some evidence anno- tations in FinLongDocQA were incorrect. We therefore re- stricted evaluation to the dataset’s table category, where ques- tions are explicitly designed to be answered from tables. The annotations were subsequently corrected by the dataset cre- ators, and we use these for all reported results. ModelTokenization Context Reasoning Gemma 4 E4BFixed128k✓ Gemma 4 26B-A4BFixed256k✓ Qwen 3 VL-8B InstructVariable256k✗ Qwen 3 VL-8B ThinkingVariable256k✓ Qwen 3.5-9BVariable256k✓ Table 2: The five open-weight VLMS we evaluate. To- kenisation 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. sentation. Tables are rendered in black and white using a uniform typeface. 3 3.3 Models Fixed-length Visual Tokenization Gemma 4 (Gemma Team, 2026) represents each image us- ing a fixed number of visual tokens selected from a discrete grid: 64, 121, 256, 529, or 1,089 tokens, independent of the image’s dimensions. We eval- uate Gemma 4 E4B and Gemma 4 26B-A4B, with context windows of 128k and 256k tokens, respec- tively. The latter is a mixture-of-experts model with 4B active parameters (see Table 2). Variable-length Visual Tokenization Qwen 3 VL (Bai et al., 2025) and Qwen 3.5 represent im- ages with a number of visual tokens that scales with their pixel area. Qwen 3 VL pairs a Qwen 3 back- bone with a SigLIP 2 encoder (Tschannen et al., 2025) continued-trained at dynamic input resolu- tions; Qwen 3.5 is natively multimodal, pre-trained end-to-end on text interleaved with visual data, but retains an encoder of the same family. In both mod- els, a visual token corresponds to a32× 32pixel region, determined by the encoder’s patch size (16) and spatial merge factor (2); so an image of size h× wyields⌈h/32⌉×⌈w/32⌉tokens. We evalu- ate the Instruct and Thinking variants of Qwen 3 VL-8B and Qwen 3.5-9B. All three configurations support 256k-token contexts. Table 2 summarizes our models; for the precise mapping from image di- mensions to visual-token budgets, see Section 3.4. 3.4 Visual Token Budgets To compare compression across visual encoders, we define five nominal visual-token budgets for each table image:B ∈64, 128, 256, 512, 1,024. Each budget is applied independently to every ta- ble image. For the Qwen models, an image of size 3 Table images were rendered using Firefox in headless mode (version 142.0.1, GeckoDriver 0.36.0) at an effective density of 96 pixels per inch. h× wproduces⌈h/32⌉×⌈w/32⌉visual tokens. If this exceedsB, we downscale the image, preserv- ing its aspect ratio, to the largest resolution whose token count does not exceed the budget; images already below the budget are not upscaled. For Gemma 4, which supports only a discrete set of visual-token configurations, we map the nominal budgets to 64, 121, 256, 529, and 1,089 tokens, respectively. All reported token counts use the ac- tual number of visual tokens entering the language decoder rather than the nominal budget. We measure efficiency by the total number of tokens processed by the language decoder: textual input tokens, visual tokens, and generated tokens. Counting both input and output captures not only the savings from compressing tables but also any additional test-time reasoning induced by compres- sion. Decoder-token count provides a common measure across model families, although it does not include the cost of visual encoding or directly measure latency. 4 Does Pixel-Level Compression Affect Table Comprehension? We examine how pixel-level compression affects two capabilities required for table question answer- ing: fine-grained reading and downstream reason- ing. We first use transcription to test whether mod- els can recover table structure and individual cell values, and then evaluate whether compressed ta- bles remain useful for answering numerical ques- tions. For each configuration, we plot task per- formance against the median number of input or generated tokens across examples. Later sections combine these quantities into total token cost. 4.1 Table Transcription Table transcription tests whether the model can read every detail of a table, including its structure and all cell values, without requiring reasoning beyond for- mat conversion. Models receive the full document, including its multiple tables and surrounding text, and are asked to transcribe every table into stan- dard HTML. We evaluate the similarity between generated and reference tables using TEDS (Zhong et al., 2020), a similarity metric based on tree-edit distance. Tables are represented as trees follow- ing their HTML structure, and TEDS is computed as one minus the tree-edit distance between the generated and reference trees, normalized by the larger tree’s size, giving a score in[0, 1]that jointly 02000400060008000 Input Total Tokens 0 20 40 60 80 100 TEDS 1024 128 128 256 256 512 512 64 64 HTML Gemma 4-26B Qwen 3.5-9B LaTeX Gemma 4-E4B Qwen 3 VL-8B Figure 2: Table transcription performance against input tokens on MultiHiertt. The HTML→HTML control is nearly perfect, while LATEX→HTML shows that format conversion introduces error even without visual percep- tion. Transcription from images degrades progressively as the visual-token budget decreases. reflects structural and cell-content accuracy. We compare three settings. In HTML→HTML, models copy the original HTML, providing a control for the generation task itself.In LA- TEX→HTML, models convert between two textual representations, allowing us to measure the diffi- culty of format conversion without visual percep- tion. Finally, IMG→HTML evaluates transcription from table images across the five visual-token bud- gets. An example is provided in Appendix A.1. Figure 2 shows TEDS against input tokens. The HTML→HTML control is nearly perfect, confirming that copying is not the bottleneck. LATEX→HTML already incurs a measurable drop, showing that part of the transcription error arises from format con- version rather than visual perception. Transcription from images remains strong at larger budgets but degrades steadily as compression increases, indicat- ing that increasingly compressed representations lose fine-grained information about table structure and cell values. For Gemma 4, the largest visual-token setting also substantially increases input cost without im- proving over the 512-token budget. This illustrates that allocating more visual tokens does not neces- sarily improve transcription once sufficient resolu- tion has been reached. 4.2 Question Answering Reduced transcription fidelity does not necessarily imply lower downstream performance: answering a question may require reading only a small portion of a table rather than reconstructing it in full. We therefore evaluate both input settings (Section 3.2) on the MultiHiertt and FinLongDocQA dataset. 2000250030003500400045005000 Input Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 128 256 512 64 HTML 1024 128 256 512 2000400060008000100001200014000 Total Tokens 0 10 20 30 40 50 60 Accuracy 128 512 64 HTML HTML 128 256 512 64 1024 256 • Gemma 4-26B• Qwen 3.5-9B Figure 3: QA accuracy on MultiHiertt (test-time rea- soning enabled) for Gemma 4 26B-A4B and Qwen 3.5- 9B, plotted against median input tokens (left) and gen- erated 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. In this analysis, the model receives the complete document; contexts assembled by an external re- triever are considered in Section 5.3. Because both datasets emphasize numerical reasoning, we enable test-time reasoning in the main experiments and report results without it in Appendix D. Figure 3 reports QA accuracy against input to- kens and generated tokens on MultiHiertt for the latest and largest model in each family. Results for all models and for FinLongDocQA are provided in Appendix E. At sufficiently large visual-token budgets, tables represented as images match and sometimes exceed the accuracy of their HTML se- rialization while using fewer input tokens. Below this point, however, accuracy generally falls as the token budget tightens. This pattern is consistent with the transcription results: once compression removes the detail needed to read cell values reli- ably, models can no longer perform the numerical operations required to answer the question. Generated tokens reveal a second, less visible cost of compression: lower visual-token budgets produce longer reasoning traces, without improv- ing performance. Inspection of the traces shows repeated misreadings of cell values and uncertainty over illegible entries, causing models to deliberate for longer. This pattern also holds for FinLong- DocQA, although its much larger number of tables means that input-token savings still outweigh the additional generation. Even so, measuring input tokens alone overstates the efficiency gains of ag- gressive compression. Full results for both datasets and all models are reported in Appendix E. Taken together, the results expose an important 641282565121024HTMLTABLEID 0 20 40 60 80 100 37.039.141.042.041.240.729.9 Gemma 4-26B r@1 random 641282565121024HTMLTABLEID 40.540.841.240.841.139.932.7 Qwen 3.5-9B F1 achievedF1 shortfall (requested)didn't request Figure 4: Relevant-table identification (F1) for Multi- Hiertt across token budgets and table representations (Gemma 4-26B, Qwen 3.5-9B). Bars (y axis), decom- pose 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 op- erating 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). limitation of direct QA over compressed tables. Moderate compression can provide useful input savings without sacrificing accuracy, but aggressive compression removes the fine-grained information needed for reasoning and induces additional gener- ation that offsets part of those savings. This raises a different possibility: heavily compressed tables may still preserve enough coarse information to identify which tables are relevant, even when they are no longer legible enough to answer from di- rectly. We test this possibility in the next section. 5 Our Two-Stage Method: Identify first, Reason Later We test whether low-resolution table images retain enough information for a model to identify the tables required to answer a question. Our two-step procedure is illustrated in Figure 1. In the first step, the model receives the document text together with all tables compressed to a fixed visual-token budget. Each table is preceded by a[TABLE_n]marker, and the model identifies the tables it requires using "request_table": n. In the second step, the requested tables are appended to the conversation at native resolution, after which the model reasons and produces the answer. The requested tables are returned at 1,024 visual tokens, which amounts to no compression on our datasets (we ablate this choice for Qwen models in Appendix G). Only the tables change between steps; the document text is always provided as text. Our prompts are provided in Appendix A. The identification step is performed without test- time reasoning, minimising its token overhead; in preliminary experiments, enabling reasoning pro- duced no significant improvement. The resulting method is training-free, prompt-level, and model- agnostic. It builds on evidence that separating iden- tification from reasoning benefits long-context QA (Guan et al., 2026; Zhao et al., 2026), while using pixel compression to make identification over the full set of document tables inexpensive. We include two controls to isolate the informa- tion supplied by the compressed images. The HTML condition applies the same two-step procedure with tables represented as text. In TABLEID, each table is replaced by its identifier, removing its contents while preserving its position and surrounding con- text. Performance above TABLEID therefore indi- cates that compressed table images provide useful evidence-identification signals. 5.1 Stage 1: Relevant-Table Identification We first evaluate evidence identification indepen- dently of downstream reasoning, using F1 over the requested tables. Our main analysis uses Multi- Hiertt, which provides direct table-level evidence annotations. FinLongDocQA instead annotates evi- dence pages, so treating every table on an evidence page as relevant produces only an upper bound on the true evidence set (we nevertheless report these results in Appendix F). Figure 4 reports results for Gemma 4 26B-A4B and Qwen 3.5-9B; results for all models are provided in Appendix F. A random- selection baseline achieves 27.4% F1. Table identification is robust to compression. Identification performance varies little across visual-token budgets: tables compressed to 64 to- kens are identified nearly as accurately as tables at higher resolutions or in HTML form. This con- trasts with transcription and direct QA, both of which degrade at the same budget (Figures 2 and 3). Aggressive compression therefore removes the fine- grained information needed to read cell values but preserves the broader cues that reveal what a table is about and whether it is relevant to the question. Compressed images provide useful evidence. The TABLEID control replaces every table with its identifier, preserving its position and surrounding text while removing its contents. Adding64-token images improves F1 by 16 for Gemma 4 26B-A4B and 10 points for Qwen 3.5-9B over this control, confirming that the model uses information re- tained in the compressed renderings rather than Full-contextBGE-M3 @5 Qwen 3.5-9BGemma 4-26BQwen 3.5-9BGemma 4-26B MultiHiertt Accuracy 600075009000105001200013500 Total Tokens 0 10 20 30 40 50 60 HTML 1024 128 256 512 64 HTML 1024 128 512 64 600075009000105001200013500 Total Tokens 0 10 20 30 40 50 60 HTML 128 256 512 64 HTML 128 256 512 64 2000300040005000600070008000 Total Tokens 0 10 20 30 40 50 60 @100 @1 @25 @2 @3 @50 1024 128 64 1024 128 256 512 64 512 256 @5 20004000600080001000012000 Total Tokens 0 10 20 30 40 50 60 @100 @1 @25 @50 @5 1024 128 256 512 64 1024 128 64 256 512 @ 3 @ 2 FinLongDocQA Accuracy 700008000090000100000110000120000 Total Tokens 0 10 20 30 40 50 60 HTML 1024 128 256 512 64 HTML 1024 128 256512 64 700008000090000100000110000120000 Total Tokens 0 10 20 30 40 50 60 HTML 128 256 512 64 HTML 128 256 512 64 0100002000030000400005000060000 Total Tokens 0 10 20 30 40 50 60 @100 @10 @200 1024 128 256 512 64 1024 128 256 512 64 0150003000045000600007500090000 Total Tokens 0 10 20 30 40 50 60 @100 @10 @200 1024 128 256512 64 1024 128 256 512 64 • Direct QA • Two-Stage QA (ours) • BGE-M3 @k Figure 5: Accuracy against total tokens for direct QA and our two-step method, on MultiHiertt (top) and FinLong- DocQA (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 increasingkas reference. Markers along each line correspond to visual token budgets, and bands show±1standard error of the mean across examples (Appendix B). Results for all models are provide in Appendix E. relying only on positional or textual cues. The TABLEID condition nevertheless remains above chance because surrounding passages often intro- duce or describe the adjacent tables. 5.2 Stage 2: Full-Resolution Reasoning Figure 5 (left panel) reports accuracy against to- tal decoder tokens, and Table 3 summarises the token savings of our method at budget 64 against single-step references: the cheapest configuration it matches in accuracy (Matched) and single-step QA over native-resolution images (Image). The Retrieval column is discussed in Section 5.3. All configurations include a fixed cost of2,655±28to- kens on MultiHiertt and61,261±348on FinLong- DocQA for the encoding of the surrounding docu- ment text. We include this cost in all reported totals but subtract it from the plot axes for readability. On MultiHiertt, the two-step procedure avoids much of the reasoning-token inflation observed in Section 4.2 at budgets of 512 tokens or fewer. The largest gain is for Gemma 4 26B-A4B: at 256 to- kens, our method uses 33.8% fewer tokens than single-step QA over native-resolution images, with a loss of 3.9 accuracy points. It is also consistently more efficient than representing tables as HTML; for Qwen 3.5-9B, it matches the accuracy of native- resolution images while using 20.9% fewer to- kens. At budget 64, our method matches the cheap- est equivalent single-step configuration while sav- ing 6.1% of tokens on average, and trails native- resolution images by only 4.4 accuracy points at 28.7% lower cost (Table 3). On FinLongDocQA, the method improves accu- racy at every budget. Its gains at 1,024 tokens show that separating evidence identification from reason- ing helps even without compression; reducing the budget then adds further efficiency while degrading much less than single-step QA. At 64 tokens, our method exceeds native-resolution single-step QA by 3.0 accuracy points while using 53.4% fewer tokens for Gemma 4 26B-A4B, and by 11.2 points while using 27.7% fewer tokens for Qwen 3.5-9B. On average across both models, it gains 7.1 ac- curacy points over native-resolution images while using 40.6% fewer tokens (Table 3). 5.3 Relationship to External Retrieval In this section we consider the relationship between our method and external retrieval. Specifically, we examine whether a dedicated retriever can re- place compressed relevant-table identification, and Identify first (64)MatchedImageRetrieval Reason later vs.Tok.Tok.Acc.Tok.Acc. MultiHiertt Qwen 3.5-9B14.020.90.025.2−0.7 Qwen 3 VL-8B7.814.8−7.817.3−2.7 Gemma 4 26B 8.843.8−8.035.9−8.0 Gemma 4 E4B−6.235.2−1.824.1−3.4 Average6.128.7−4.425.6−3.7 FinLongDocQA Qwen 3.5-9B11.127.7+11.251.0+2.2 Gemma 4 26B 18.753.4+3.072.5+0.2 Average14.940.6+7.161.8+1.2 Table 3: Percentage of tokens saved (Tok.) and ac- curacy 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. whether our method remains useful after retrieval has already reduced the context. All retrievers are used zero-shot with their default pretrained weights and the original question as the query; tables are treated as atomic units. External retrieval identifies better but costs more. We replace the identification stage with ColQwen2 (Faysse et al., 2024), a multi-vector re- triever designed for visually represented text and the stronger retriever in our table-retrieval experi- ments (Appendix H). Retrieved tables are supplied uncompressed alongside the document text, and we evaluate several values of k on both datasets. ColQwen2 identifies evidence tables more ac- curately than models reading a compressed con- text (Figure 4), but this advantage does not carry through to downstream accuracy. On MultiHiertt, retrieval plateaus at the accuracy our method al- ready reaches: for Qwen 3.5-9B, our 64-token con- figuration matches the accuracy of the cheapest most accurate ColQwen2 configuration, using 19% fewer tokens. Only for Gemma 4 26B-A4B does retrieval overtake our method, but never by more than 7.3 accuracy points, which our method reaches using 45.3% of the tokens. On FinLongDocQA our method exceeds retrieval by 16 accuracy points at the same cost. External retrieval therefore identi- fies better without delivering a better end-to-end trade-off. Full curves are in Appendix I. Our two-step method still helps after retrieval. We next apply our two-step method to contexts re- trieved by BGE-M3 (Chen et al., 2024a) from a unified pool of paragraphs and tables. We usek=5 on MultiHiertt andk=100on FinLongDocQA, ren- der the retrieved tables at each visual-token budget, and compare against single-step QA over the same retrieved context (see Figure 5 right panel). On MultiHiertt, retrieval performance falls sharply belowk=5, limiting how far retrieval alone can reduce the context. Our method moves beyond this limit: Qwen 3.5-9B matches the accuracy of single-step QA atk=5using 25.2% less tokens, while Gemma 4 26B-A4B at a budget of 256 uses 26.6% fewer tokens at a cost of 2.9 accuracy points. The same general pattern holds for the remaining models except Gemma 4 E4B. As in Section 5.2, compression alone does not achieve these gains; they arise from identifying and restoring only the tables needed for reasoning. On FinLongDocQA, our method matches the accuracy of retrieval using 51.0% and 72.5% fewer tokens respectively (see Table 3, Retrieval column). 6 Conclusions We investigated whether pixel-level compression can reduce the cost of question answering over multi-table documents. Moderate compression al- lows table images to match their HTML serializa- tion while using fewer tokens. Under aggressive compression, however, models lose the cell-level information needed for numerical reasoning and reason for longer, eroding the efficiency gains. Our central finding is that compression affects reading and identification differently. Even when a table is too compressed to support accurate reason- ing, it retains enough information for identifying whether it is relevant. We exploit this asymmetry with a training-free, two-step method that first iden- tifies relevant tables in a compressed context and then restores only those tables for full-resolution reasoning. The method improves the accuracy- efficiency trade-off in full-context QA and on top of retrieval, using 41% fewer tokens while gain- ing 7 accuracy points over single-step QA with native-resolution tables on long documents. More broadly, our results suggest that pixel com- pression is most effective not as a representation from which models must reason over directly, but as an inexpensive intermediate step for locating evidence before native-resolution computation. Limitations Our experiments are limited to English-language financial documents and two datasets; moreover, MultiHiertt results are reported on the development set for the reasons discussed in Section 3.1. We render tables from clean structured representations using a uniform style in order to preserve informa- tion parity between the textual and visual settings. Real-world table images may contain varied fonts, complex layouts, scanning artifacts, and recogni- tion errors. Our experiments also assume that ta- ble boundaries are known; applying the method to raw documents may require an additional table- detection or extraction stage. The effective compression level depends on table dimensions, information density, rendering style, and the model’s visual encoder. The specific token budgets identified here may therefore not transfer directly to other domains or document formats, al- though the evaluation procedure can be used to determine suitable budgets in new settings. Finally, we use the number of decoder tokens as an architecture-independent proxy for efficiency. This measure captures both input and generated tokens, but not visual-encoder computation, the latency and overhead of the additional inference pass, or hardware-dependent differences in cost. Measuring wall-clock latency, memory use, and FLOPs would be necessary to establish the full computational trade-off. References Iñigo Alonso, Eneko Agirre, and Mirella Lapata. 2024. PixT3: Pixel-based table-to-text generation. In Pro- ceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 6721–6736, Bangkok, Thailand. As- sociation for Computational Linguistics. Iñigo Alonso, Imanol Miranda, Eneko Agirre, and Mirella Lapata. 2026. Tablet: A large-scale dataset for robust visual table understanding.Preprint, arXiv:2509.21205. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhi- fang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, and 45 others. 2025. Qwen3-vl technical report. Preprint, arXiv:2511.21631. Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a.M3- embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self- knowledge distillation. In Findings of the Asso- ciation for Computational Linguistics: ACL 2024, pages 2318–2335, Bangkok, Thailand. Association for Computational Linguistics. Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Jun- yang Lin, Chang Zhou, and Baobao Chang. 2024b. An image is worth 1/2 tokens after layer 2: Plug-and- play inference acceleration for large vision-language models. Preprint, arXiv:2403.06764. Tong Chen, Hao Fang, Patrick Xia, Xiaodong Liu, Ben- jamin Van Durme, Luke Zettlemoyer, Jianfeng Gao, and Hao Cheng. 2024c. Generative adapter: Con- textualizing language models in parameters with a single forward pass. Preprint, arXiv:2411.05877. Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. Tabfact: A large-scale dataset for table-based fact verification. In Inter- national Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia. Jiale Cheng, Yusen Liu, Xinyu Zhang, Yulin Fei, Wenyi Hong, Ruiliang Lyu, Weihan Wang, Zhe Su, Xiaotao Gu, Xiao Liu, Yushi Bai, Jie Tang, Hongning Wang, and Minlie Huang. 2026. Glyph: Scaling context windows via visual-text compression. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 37145–37158, San Diego, California, United States. Association for Computational Linguistics. Joonmyung Choi, Sanghyeok Lee, Jongha Kim, Se- hyung Kim, Dohwan Ko, Jihyung Kil, and Hyun- woo J. Kim. 2026. Docprune:efficient document question answering via background, question, and comprehension-aware token pruning.Preprint, arXiv:2604.22281. Naihao Deng, Zhenjie Sun, Ruiqi He, Aman Sikka, Yu- long Chen, Lin Ma, Yue Zhang, and Rada Mihalcea. 2024. Tables as texts or images: Evaluating the table reasoning ability of LLMs and MLLMs. In Find- ings of the Association for Computational Linguis- tics: ACL 2024, pages 407–426, Bangkok, Thailand. Association for Computational Linguistics. Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2024. Colpali: Efficient document retrieval with vi- sion language models. Preprint, arXiv:2407.01449. Tianyu Gao, Zirui Wang, Adithya Bhaskar, and Danqi Chen. 2024. Improving language understanding from screenshots. arXiv preprint arXiv:2402.14073. Gemma Team. 2026.Gemma 4 technical report. Preprint, arXiv:2607.02770. Xin Guan, Zijian Li, Shen Huang, Pengjun Xie, Jingren Zhou, and Jiuxin Cao. 2026. Evidence-augmented policy optimization with reward co-evolution for long-context reasoning. Preprint, arXiv:2601.10306. Michael Günther, Isabelle Mohr, Daniel James Williams, Bo Wang, and Han Xiao. 2025. Late chunking: Con- textual chunk embeddings using long-context embed- ding models. Preprint, arXiv:2409.04701. Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Müller, Francesco Piccinno, and Julian Eisenschlos. 2020. TaPas: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4320–4333, Online. Association for Computa- tional Linguistics. Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shan- tanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. 2024. Ruler: What’s the real context size of your long-context language models? arXiv preprint arXiv:2404.06654. Yoonsik Kim, Moonbin Yim, and Ka Yeon Song. 2024. Tablevqa-bench: A visual question answer- ing benchmark on multiple table domains. Preprint, arXiv:2404.19205. Tung Sum Thomas Kwok, Xinyu Wang, Xiaofeng Lin, Peng Lu, Chunhe Wang, Changlun Li, Han- wei Wu, Nan Tang, Elisa Kreiss, and Guang Cheng. 2026. Tabqaworld: Optimizing multimodal reason- ing for multi-turn table question answering. Preprint, arXiv:2604.03393. Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandel- wal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. 2023. Pix2struct: screenshot parsing as pretraining for visual language understanding. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org. Yanhong Li, Zixuan Lan, and Jiawei Zhou. 2025. Text or pixels? evaluating efficiency and understanding of LLMs with visual text inputs. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 10564–10578, Suzhou, China. Associa- tion for Computational Linguistics. Jiaheng Liu, Dawei Zhu, Zhiqi Bai, Yancheng He, Huanxuan Liao, Haoran Que, Zekun Wang, Chenchen Zhang, Ge Zhang, Jiebin Zhang, Yuanx- ing Zhang, Zhuo Chen, Hangyu Guo, Shilong Li, Ziqiang Liu, Yong Shan, Yifan Song, Jiayi Tian, Wen- hao Wu, and 18 others. 2025. A comprehensive sur- vey on long context language modeling. Preprint, arXiv:2503.17407. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language mod- els use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173. Jonas Lotz, Elizabeth Salesky, Phillip Rust, and Desmond Elliott. 2023. Text rendering strategies for pixel language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, pages 10155–10172, Singapore. Association for Computational Linguistics. Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Man- aal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipan- jan Das. 2020. ToTTo: A controlled table-to-text generation dataset. In Proceedings of the 2020 Con- ference on Empirical Methods in Natural Language Processing (EMNLP), pages 1173–1186, Online. As- sociation for Computational Linguistics. Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Eliza- beth Salesky, Miryam de Lhoneux, and Desmond Elliott. 2023.Language modelling with pixels. Preprint, arXiv:2207.06991. Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. 2025. Llava-prumerge: Adaptive to- ken reduction for efficient large multimodal models. ICCV. Anshul Singh, Chris Biemann, and Jan Strich. 2025. MTabVQA: Evaluating multi-tabular reasoning of language models in visual space. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19866–19891, Suzhou, China. Associa- tion for Computational Linguistics. Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. 2023. Table meets llm: Can large language models understand structured table data? a benchmark and empirical study. Web Search and Data Mining. Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. 2025. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. Preprint, arXiv:2502.14786. Alex Jinpeng Wang, Linjie Li, Yiqi Lin, Min Li, Lijuan Wang, and Mike Zheng Shou. 2024a. Leveraging vi- sual tokens for extended text contexts in multi-modal learning. Preprint, arXiv:2406.02547. Lanrui Wang, Mingyu Zheng, Hongyin Tang, Zheng Lin, Yanan Cao, Jingang Wang, Xunliang Cai, and Weiping Wang. 2025.Needleinatable: Ex- ploring long-context capability of large language models towards long-structured tables. Preprint, arXiv:2504.06560. Yi-Cheng Wang, Wei-An Wang, and Chu-Song Chen. 2026. Document-level numerical reasoning across single and multiple tables in financial reports. arXiv preprint arXiv:2604.03664. Zilong Wang, Hao Zhang, Chun-Liang Li, Julian Martin Eisenschlos, Vincent Perot, Zifeng Wang, Lesly Mi- culicich, Yasuhisa Fujii, Jingbo Shang, Chen-Yu Lee, and Tomas Pfister. 2024b. Chain-of-table: Evolving tables in the reasoning chain for table understanding. ICLR. Haoran Wei, Yaofeng Sun, and Yukun Li. 2025. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234. Mingyuan Wu, Jingcheng Yang, Jize Jiang, Meitang Li, Kaizhuo Yan, Hanchao Yu, Minjia Zhang, Chengxi- ang Zhai, and Klara Nahrstedt. 2025. Vtool-r1: Vlms learn to think with images via reinforcement learning on multimodal tool use. Preprint, arXiv:2505.19255. Ling Xing, Alex Jinpeng Wang, Rui Yan, Xiangbo Shu, and Jinhui Tang. 2025a. Vision-centric token com- pression in large language model. arXiv preprint arXiv:2502.00791. Xiaobo Xing, Wei Yuan, Tong Chen, Quoc Viet Hung Nguyen, Xiangliang Zhang, and Hongzhi Yin. 2025b. Tabledart: Dynamic adaptive multi-modal routing for table understanding. Preprint, arXiv:2509.14671. Chanwoong Yoon, Taewhoo Lee, Hyeon Hwang, Min- byul Jeong, and Jaewoo Kang. 2024. CompAct: Compressing retrieved documents actively for ques- tion answering. In Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Pro- cessing, pages 21424–21439, Miami, Florida, USA. Association for Computational Linguistics. Yusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister, Rui Zhang, and Sercan Ö. Arik. 2024. Chain of agents: Large language models collaborating on long- context tasks. Preprint, arXiv:2406.02818. Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, and Jingrui He. 2026. Recontext: Recursive evidence replay as llm harness for long- context reasoning. Preprint, arXiv:2607.02509. Yilun Zhao, Yunxiang Li, Chenying Li, and Rui Zhang. 2022. MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland. Association for Computational Linguistics. Mingyu Zheng, Xinwei Feng, Qingyi Si, Qiaoqiao She, Zheng Lin, Wenbin Jiang, and Weiping Wang. 2024. Multimodal table understanding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9102–9124, Bangkok, Thailand. Association for Computational Linguistics. Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. 2020. Image-based table recognition: data, model, and evaluation. Preprint, arXiv:1911.10683. Jiaru Zou, Soumya Roy, Vinay Kumar Verma, Ziyi Wang, David Wipf, Pan Lu, Sumit Negi, James Zou, and Jingrui He. 2025. Tattoo: Tool-grounded think- ing prm for test-time scaling in tabular reasoning. Preprint, arXiv:2510.06217. A Prompts This appendix shows the prompts used in our exper- iments Figures 6–8). For readability, each prompt shows a single table image, but every<image> placeholder in the actual prompts is accompanied by its corresponding table image. The tables shown here are included for reference only and are not rendered at the pixel dimensions of any particular visual-token budget. A.1 Transcription (OCR) Figure 6 shows the prompt used for table tran- scription, in the IMG→HTML setting.In the HTML→HTML and LATEX→HTML settings, each <image>placeholder is replaced by the HTML or LaTeX serialization of the same table, respectively, and the rest of the prompt remains unchanged. A.2 Question Answering (QA) Figures 7 and 8 show the prompts used for direct QA and for our two-stage method, both in the hybrid setting. The HTML setting replaces each <image>placeholder with the HTML serialization of the same table, and the rest of the prompt re- mains unchanged. B Confidence intervals Every score we report is a mean over a fixed set ofNtest examples. We take the test example as the unit of analysis and report the standard error of the mean (s.e.m.). Lets i ∈0, 1be a model’s exact-match accuracy on examplei. We report the meanˆμ = 1 N P i s i (shown as a percentage) with a band of±1s.e.m., a68%interval under a nor- mal approximation, which captures the sampling uncertainty of each plotted mean: s.e.m. = ˆσ √ N , ˆσ 2 = 1 N − 1 N X i=1 s i − ˆμ 2 . (1) C Quantifying Token Savings Each configuration in this work gives us two num- bers, accuracy and token cost. We know that the Pareto frontier lies in the upper-left corner of our accuracy per token graphs, but defining its precise gradient requires choosing which of the two to prioritize. When selecting a point of reference to compare our method against, we prioritize accu- racy. That is, we compare the cheapest and most accurate configuration of each method, first select- ing the most accurate and then, among the results ============================================= prompt ============================================= You are given a financial document containing text and interleaved tables. Your task is to convert ONLY the tables into HTML. Ignore the rest of the document. Rules: - Preserve the exact structure, hierarchy, and content of each table. - Use ONLY these tags: <td>, <tr>, <table>, <sup>, <i>, <sub>. And only the following attributes when needed: colspan, rowspan. Do not use any other HTML tags or attributes. - Do not add, omit, or modify any cell content. Reproduce it exactly as it appears. - Output one HTML string per table, in order of appearance in the document, one after the other <table>...</table><table>...</table>. Example: This table <image> should be converted as: <table><tr><td></td><td colspan=\"2\">As of December</td></tr><tr><td><i>$ in millions</i></td><td> 2015</td><td>2014</td></tr><tr><td>Collateral available to be delivered or repledged<sup>1</sup></ td><td> $636,684</td><td>$630,046</td></tr><tr><td>Collateral that was delivered or repledged</ td><td> 496,240</td><td>474,057</td></tr></table> Document: // ... text omitted in the paper ... // <image> // ... text omitted in the paper ... // <image> // ... text omitted in the paper ... // Figure 6: Prompt used for table transcription. ============================================= prompt ============================================= You are given a financial document containing text and interleaved tables. Answer the following question based ONLY on the provided document. You may reason about the answer, but you MUST include your final answer in the following JSON format exactly once in your response: "answer": "YOUR_ANSWER" Rules: - Provide the answer exactly as it appears or can be derived from the document. - Do not include units, currency symbols, or extra text inside the answer unless the question asks for them. - If the answer is a number, provide it as a plain number (e.g., \"21164\", not \"$21,164\" or \"21,164 million\"). Example: Question: What's the greatest value of Consumer in 2011? (in million) Answer: \"answer\": \"21164\" Document: // ... text omitted in the paper ... // <image> // ... text omitted in the paper ... // <image> Question: What's the total amount of the elements in the years where Corporate is greater than 70000? Figure 7: Prompt used for direct QA. ========================================= Step 1 prompt =========================================== You are given a financial document containing text and interleaved tables. In this stage you MUST request the table(s) you need in order to answer the given question. Do not answer the question and do not reason in this first step, focus only on requesting the right table to answer the question. You will reason later when the table is retrieved. Each table in the document is associated with an identifier of the form [TABLE_n], where n is its number in the document. To request a table, emit JSON in this format: "request_table": n where n is the table's number (e.g., to request [TABLE_0] emit "request_table": 0). If answering genuinely requires data from more than one table, you may emit several such objects in the same response (e.g., "request_table": 3,"request_table": 1). Request ONLY the specific table(s) strictly necessary to answer this question. The requested table(s) will then be provided, and you will answer in the next step. Document: // ... text omitted in the paper ... // [TABLE_0] <image> // ... text omitted in the paper ... // [TABLE_n] <image> Question: What's the total amount of the elements in the years where Corporate is greater than 70000? ========================================= Step 2 prompt =========================================== Below are the table(s) you requested: [TABLE_3] <image> Now answer the question using the original document together with the table(s) above. You can no longer request additional tables. You may reason about the answer, but you MUST include your final answer in the following JSON format exactly once in your response: \"answer\": \"YOUR_ANSWER\" Rules: - Provide the answer exactly as it appears or can be derived from the document. - Do not include units, currency symbols, or extra text inside the answer unless the question asks for them. - If the answer is a number, provide it as a plain number (e.g., \"21164\", not \"$21,164\" or \"21,164 million\"). Example: Question: What's the greatest value of Consumer in 2011? (in million) Answer: \"answer\": \"21164\" Question: What's the total amount of the elements in the years where Corporate is greater than 70000? Figure 8: Prompt used for our two-stage method. Algorithm 1 Token savings at matched accuracy 1:Order the baseline results by accuracy, highest first, and take each in turn as the target level. 2:Collect every configuration, from either method, whose accuracy is not below the target once error intervals are accounted for. 3:If none of our method’s configurations quali- fies, the target is out of reach; skip to the next. 4: Otherwise take the cheapest qualifying config- uration of our method and the cheapest quali- fying baseline. 5:Report the percentage of tokens the former saves over the latter. at equivalent performance, the cheapest. We de- fine equivalence using the confidence intervals of Appendix B. We also assume that the baseline can represent tables as images and that pixel-level compression is available to it as well. Note that this strips from our results the improvements that come from those two techniques alone, so what remains comes solely from the two-stage procedure itself, making this a conservative estimate and, in our view, a fairer and more realistic comparison. Algorithm 1 states the procedure. D Effect of Test-Time Reasoning Figure 9 compares direct QA with and without test- time reasoning across visual token budgets. Two patterns emerge. First, the token cost of reasoning grows as tables are compressed, so the gap between the two conditions widens towards the lower bud- gets, which is the effect discussed in Section 4.2. Second, the accuracy benefit of reasoning varies by model, but is predominantly positive across all models. We enable test-time reasoning in the main experiments because both datasets emphasize nu- merical reasoning, and because it is the stronger of the two conditions for most models. E Full Results This appendix reports the complete results for Sec- tions 4.2, 5.2, and 5.3, for all models and both datasets. Figures show accuracy against total to- kens; tables give the underlying numbers, including the breakdown of token counts into their textual, vi- sual, and generated tokens. On FinLongDocQA we evaluate only the largest and most recent model of each family, as Gemma 4 E4B’s 128k context win- dow cannot accommodate these documents along- side reasoning traces, and each configuration is considerably more expensive to run at this length. E.1 Full-Context QA In this setting, the complete document is provided to the model, with tables represented as images at each visual token budget (Figure 10). Direct QA answers from this context in a single step, while our two-stage method first identifies the tables it needs and then reasons over them at native resolution. E.2 QA over Retrieved Context In this setting, the context is assembled by BGE- M3 rather than fed in full, usingk=5on Multi- Hiertt and k=100 on FinLongDocQA (Figure 11). Direct QA and our two-stage method operate over this reduced context, with single-step QA over re- trieved units at increasingkincluded as a reference for what retrieval alone achieves. F Table Identification: All Models Figures 12 and 13 extend the identification re- sults of Section 5.1 to every model. Robustness to compression holds across the board: identifica- tion varies little across visual token budgets for all models. The gain over the TABLEID is clear for Qwen 3.5 and the Gemma 4 models, but marginal for Qwen 3 VL. On FinLongDocQA, identification F1 is almost identical between low image budgets and TABLEID. As noted in Section 5.1, this dataset annotates evi- dence at the page level, so treating every table on an evidence page as relevant yields an upper bound on the true evidence set. The downstream accuracies printed above each bar tell a different story. Remov- ing table contents costs Qwen 3.5 a large amount of accuracy (37.0 at 64 tokens against 17.8 for TABLEID) and Gemma 4 26B-A4B a smaller one (43.2 against 41.2), even though the two conditions are indistinguishable under F1. This is consistent with the annotation being too coarse to separate them: a model that locates the right page scores well whether or not it can tell which of its tables matters, while the answer still depends on getting that distinction right. G At What Resolution to Return the Requested Table The two stages of our method have independent budgets: the first compresses every table in the con- Qwen 3 VLQwen 3.5-9BGemma 4-E4BGemma 4-26B 010000200003000040000 Total Tokens 0 20 40 60 80 100 Accuracy 1024 1024 128 128 256 256 512 512 64 64 HTML HTML 0200040006000800010000 Total Tokens 0 20 40 60 80 100 Accuracy 1024 1024 128 128 256 256 512 512 64 64 HTML HTML 02000400060008000 Total Tokens 0 20 40 60 80 100 Accuracy 1024 1024 128 128 256 256 512 512 64 64 HTML HTML 02000400060008000100001200014000 Total Tokens 0 20 40 60 80 100 Accuracy 1024 1024 128 128 256 256 512 512 64 64 HTML HTML • Without test-time reasoning • With test-time reasoning 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. text, while the second returns the requested table at a chosen resolution. The main text fixes the re- turn budget at 1,024 visual tokens, which amounts to no compression on our datasets. Figure 14 jus- tifies this choice on Qwen 3.5-9B, sweeping the return budget over HTML and the five visual-token budgets for each of the five context budgets. Accuracy holds up as long as the returned table is legible: HTML and 1,024 visual tokens perform equally, and accuracy begins to degrade from 512 tokens onwards—slightly at first, then more clearly at the lower budgets. Generated tokens mirror the finding of Section 4.2: returning the evidence at a low budget reintroduces the longer reasoning traces that compression induces, so the second stage loses the very property that motivates it. We therefore re- turn the requested table at 1,024 tokens, which pre- serves accuracy and keeps reasoning traces short, and which on our datasets is equivalent to returning it uncompressed. H Retriever Comparison This appendix compares the two retrievers used in the paper and justifies where each is applied. BGE-M3 (Chen et al., 2024a) indexes linearised text, while ColQwen2 (Faysse et al., 2024) indexes rendered images, so only the former can operate over a pool that mixes paragraphs and tables. Fig- ure 15 reports average recall and F1 against the number of retrieved units, under three settings. In any (paragraph) and any (table), the pool contains both the document’s paragraphs and its tables, and we measure how well the retriever recovers para- graph evidence and table evidence respectively; only BGE-M3 applies here. In tables, the pool con- tains tables alone, which allows both retrievers to be compared directly, BGE-M3 over their textual serialization and ColQwen2 over their rendered images. ColQwen2 outperforms BGE-M3 at retrieving tables from a table-only pool, which is why we use it in the table-retrieval setting of Section 5.3 and as the reference operating point in the identification results of Section 5.1. BGE-M3 is used wherever the pool mixes modalities, since ColQwen2 cannot index text. On FinLongDocQA we omit the any (para- graph) setting, as the dataset annotates evidence at the page level rather than per passage. For the same reason, we treat every table on an evidence page as an evidence table, which inflates the size of the evi- dence set and is part of why recall approaches 1.0 more slowly than on MultiHiertt; the far larger can- didate pool, around 80 tables per document against three, accounts for the rest. I Retriever for Identification Section 5.3 asks whether a dedicated retriever can replace the identification stage of our method. Fig- ure 16 reports the full curves behind that compari- son, for every model and both datasets. In the re- trieval condition the document text is fed in full, as in our hybrid setting, but the tables are supplied by ColQwen2 at increasingkand left uncompressed, so the retriever assumes the role our first stage oth- erwise plays. MultiHierttFinLongDocQA Qwen 3 VLQwen 3.5-9BQwen 3.5-9B 010000200003000040000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 1024 128 256 512 64 HTML 1024 128 256 64 0200040006000800010000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 1024 128 256 512 64 HTML 1024 128 256 512 64 020000400006000080000100000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 1024 128 256 512 64 HTML 1024 128 256512 64 Gemma 4-E4BGemma 4-26BGemma 4-26B 02000400060008000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML1024 128 256 512 64 HTML 1024 128 256 512 64 02000400060008000100001200014000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 1024 128 256 512 64 HTML 1024 128 256 512 64 0300006000090000120000150000 Total Tokens 0 10 20 30 40 50 60 Accuracy HTML 1024 128 256 512 64 HTML 128 256 512 64 • Direct QA • Two-Stage QA (ours) 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±1s.e.m. across examples (Appendix B). MultiHierttFinLongDocQA Qwen 3 VLQwen 3.5-9BQwen 3.5-9B 01000200030004000500060007000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @1 @25 @2 @3 @50 @51024 128 256 512 64 128 256 512 64 02000400060008000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @1 @25 @2 @3 @50 @51024 128 256 512 64 1024 128 256 512 64 010000200003000040000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @10 @200 1024 128 256 512 64 1024 128 256 512 64 Gemma 4-E4BGemma 4-26BGemma 4-26B 010002000300040005000600070008000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @1 @25 @2 @3 @50 @51024 128 256 512 64 1024 128 256 512 64 020004000600080001000012000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @1 @25 @2 @3 @50 @5 1024 128 256 512 64 1024 128 256 512 64 020000400006000080000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @10 @200 1024 128 256512 64 1024 128 256 512 64 • Direct QA • Two-Stage QA (ours) • BGE-M3 @k Figure 11: Accuracy against total tokens for direct QA, our two-step method, on MultiHiertt (top) and FinLong- DocQA (bottom), for Qwen 3 VL 8B, Qwen 3.5 9B, Gemma 4 E4B, and Gemma 4 26B-A4B. We feed a context retrieved by BGE-M3, with single-step QA over the retrieved units at increasingkas reference. Markers along each line correspond to visual token budgets, and bands show±1 s.e.m. across examples (Appendix B). 641282565121024HTMLTABLEID 0 20 40 60 80 100 34.536.639.039.239.737.434.6 Qwen 3 VL-8B r@1 random 641282565121024HTMLTABLEID 40.540.841.240.841.139.932.7 Qwen 3.5-9B 641282565121024HTMLTABLEID 37.039.141.042.041.240.729.9 Gemma 4-26B 641282565121024HTMLTABLEID 33.735.135.236.235.736.326.4 Gemma 4-E4B F1 achievedF1 shortfall (requested)didn't request 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. 641282565121024HTMLTABLEID 0 20 40 60 80 100 37.040.242.242.242.838.217.8 Qwen 3.5-9B r@4 random 641282565121024HTMLTABLEID 43.245.847.246.241.550.241.2 Gemma 4-26B F1 achievedF1 shortfall (requested)didn't request Figure 13: Relevant-table identification (F1) on Fin- LongDocQA 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 anno- tates 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. 30003500400045005000550060006500 Input Total Tokens 0 10 20 30 40 50 60 Accuracy 1024 1024 1024 1024 1024 512 512 512 512 256 256 128 64 HTML HTML HTML HTML HTML 1024 128 256 512 64 30004000500060007000800090001000011000 Total Tokens 0 10 20 30 40 50 60 Accuracy 1024 1024 1024 512 512 256 128 64 HTML HTML HTML HTML 1024 128 256 512 64 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). 11325374961738597109 0.0 0.2 0.4 0.6 0.8 1.0 mean recall@k any (paragraph) 11325374961738597109 any (table) 1234567 tables 11325374961738597109 k 0.0 0.2 0.4 0.6 0.8 1.0 mean f1@k 11325374961738597109 k 1234567 k textrandomimage 51953855757659551145133515251715 0.0 0.2 0.4 0.6 0.8 1.0 mean recall@k any (table) 114274053667992105118 tables 51953855757659551145133515251715 k 0.0 0.2 0.4 0.6 0.8 1.0 mean f1@k 114274053667992105118 k textrandomimage 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. MultiHierttFinLongDocQA Qwen 3 VLQwen 3.5-9BQwen 3.5-9B 5000550060006500700075008000 Total Tokens 0 10 20 30 40 50 60 Accuracy @1 @2 @3 @5 1024 128 256 512 64 5000550060006500700075008000 Total Tokens 0 10 20 30 40 50 60 Accuracy @1 @2 @3@51024 128 256 512 64 60000700008000090000100000110000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @10 @2 @4 @50 @75 1024 128 256512 64 Gemma 4-E4BGemma 4-26BGemma 4-26B 400050006000700080009000 Total Tokens 0 10 20 30 40 50 60 Accuracy @1 @2 @3 @5 1024 128 256 512 64 50006000700080009000100001100012000 Total Tokens 0 10 20 30 40 50 60 Accuracy @1 @2 @3 @5 1024 128 256 512 64 6000080000100000120000140000160000 Total Tokens 0 10 20 30 40 50 60 Accuracy @100 @10 @2 @4 @50 @75 1024 128 256 512 64 • Two-Stage QA with model-based evidence identification (ours) • ColQwen2 @k 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 increasingk, 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±1s.e.m. across examples (Appendix B). ModelBudget TaskAcc. Input text Input visual Output textTotal Qwen 3 VL-8B-Dir. QA24.84696032768 36752 Qwen 3 VL-8B1024Dir. QA42.22601135127456990 Qwen 3 VL-8B512Dir. QA40.02601108925636682 Qwen 3 VL-8B256Dir. QA34.5260172231006574 Qwen 3 VL-8B128Dir. QA28.2260143035267053 Qwen 3 VL-8B64Dir. QA17.8260129445227346 Qwen 3 VL-8B-2Stage QA 37.45987039208094 Qwen 3 VL-8B10242Stage QA 39.72884192334576812 Qwen 3 VL-8B5122Stage QA 39.22886174236886435 Qwen 3 VL-8B2562Stage QA 39.02886134439826162 Qwen 3 VL-8B1282Stage QA 36.6287896649915901 Qwen 3 VL-8B642Stage QA 34.5287276659905956 Qwen 3.5-9B-Dir. QA39.64765049589946 Qwen 3.5-9B1024Dir. QA40.52650135829747472 Qwen 3.5-9B512Dir. QA39.12650107028186874 Qwen 3.5-9B256Dir. QA37.3265071731406780 Qwen 3.5-9B128Dir. QA29.8265042538047206 Qwen 3.5-9B64Dir. QA23.2265029252388466 Qwen 3.5-9B-2Stage QA 39.96064021908795 Qwen 3.5-9B10242Stage QA 41.12959189117727008 Qwen 3.5-9B5122Stage QA 40.82962171616956536 Qwen 3.5-9B2562Stage QA 41.22960133818246282 Qwen 3.5-9B1282Stage QA 40.8296098418826110 Qwen 3.5-9B642Stage QA 40.5295982218685910 Gemma 4-E4B-Dir. QA35.5475607905723 Gemma 4-E4B1024Dir. QA35.5263242926887422 Gemma 4-E4B512Dir. QA35.2263221147585498 Gemma 4-E4B256Dir. QA32.9263210287864528 Gemma 4-E4B128Dir. QA29.926324948394115 Gemma 4-E4B64Dir. QA23.126322359023858 Gemma 4-E4B-2Stage QA 36.3597805606679 Gemma 4-E4B10242Stage QA 35.7384142925748590 Gemma 4-E4B5122Stage QA 36.2383921145576600 Gemma 4-E4B2562Stage QA 35.2383010285665566 Gemma 4-E4B1282Stage QA 35.138204945985034 Gemma 4-E4B642Stage QA 33.738162356154810 Gemma 4-26B A4B -Dir. QA44.24756043329666 Gemma 4-26B A4B 1024Dir. QA44.9263242923874 11326 Gemma 4-26B A4B 512Dir. QA43.72632211441129280 Gemma 4-26B A4B 256Dir. QA40.82632102851389118 Gemma 4-26B A4B 128Dir. QA36.4263249467399982 Gemma 4-26B A4B 64Dir. QA24.7263223510288 13339 Gemma 4-26B A4B -2Stage QA 40.75983018698508 Gemma 4-26B A4B 10242Stage QA 41.2388642921956 10540 Gemma 4-26B A4B 5122Stage QA 42.03886211419688468 Gemma 4-26B A4B 2562Stage QA 41.03870102822307492 Gemma 4-26B A4B 1282Stage QA 39.1385249421386948 Gemma 4-26B A4B 642Stage QA 37.0377823519486360 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. ModelBudget TaskAcc. Input text Input visual Output textTotal Qwen 3.5-9B-Dir. QA29.010278204760 108016 Qwen 3.5-9B1024Dir. QA25.861050334443690 100936 Qwen 3.5-9B512Dir. QA25.56105024916384491947 Qwen 3.5-9B256Dir. QA25.06105015950411682064 Qwen 3.5-9B128Dir. QA19.0610508914500675332 Qwen 3.5-9B64Dir. QA14.8610506064525273452 Qwen 3.5-9B-2Stage QA 38.210422203949 108494 Qwen 3.5-9B10242Stage QA 42.861973337322342 101184 Qwen 3.5-9B5122Stage QA 42.26197425564243491910 Qwen 3.5-9B2562Stage QA 42.26197416652233682077 Qwen 3.5-9B1282Stage QA 40.2619749398240475324 Qwen 3.5-9B642Stage QA 37.0619746562292372930 Gemma 4-26B A4B -Dir. QA44.510256204984 109496 Gemma 4-26B A4B 1024Dir. QA40.260829858924642 157397 Gemma 4-26B A4B 512Dir. QA40.860829423405398 112222 Gemma 4-26B A4B 256Dir. QA40.06082920630568690110 Gemma 4-26B A4B 128Dir. QA38.0608299990671280083 Gemma 4-26B A4B 64Dir. QA28.7608294814818275411 Gemma 4-26B A4B -2Stage QA 50.210415402796 109296 Gemma 4-26B A4B 10242Stage QA 41.562219858923154 158804 Gemma 4-26B A4B 5122Stage QA 46.262948423403088 112073 Gemma 4-26B A4B 2562Stage QA 47.26246820630336889158 Gemma 4-26B A4B 1282Stage QA 45.8627649990324878417 Gemma 4-26B A4B 642Stage QA 43.2626964814344473280 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. ModelBudgetTaskAcc. Input text Input visual Output textTotal Qwen 3 VL-8B1024@100 Re. QA 39.02583135828287005 Qwen 3 VL-8B1024@50Re. QA 39.41927134829016372 Qwen 3 VL-8B1024@25Re. QA 39.51091126026795153 Qwen 3 VL-8B1024@5Re. QA 37.535876829344066 Qwen 3 VL-8B1024@3Re. QA 36.929061629523936 Qwen 3 VL-8B1024@2Re. QA 32.825553131883940 Qwen 3 VL-8B1024@1Re. QA 28.722133633183858 Qwen 3.5-9B1024@100 Re. QA 41.22650135828897184 Qwen 3.5-9B1024@50Re. QA 40.31978134829916502 Qwen 3.5-9B1024@25Re. QA 40.71122126032015668 Qwen 3.5-9B1024@5Re. QA 39.537276833414462 Qwen 3.5-9B1024@3Re. QA 36.730261637754630 Qwen 3.5-9B1024@2Re. QA 34.226753142084896 Qwen 3.5-9B1024@1Re. QA 28.623233643184818 Gemma 4-E4B1024@100 Re. QA 35.4263242927447498 Gemma 4-E4B1024@50Re. QA 35.9197942847286798 Gemma 4-E4B1024@25Re. QA 36.0112632767585637 Gemma 4-E4B1024@5Re. QA 34.738121447192950 Gemma 4-E4B1024@3Re. QA 33.031111006862514 Gemma 4-E4B1024@2Re. QA 30.027610856712114 Gemma 4-E4B1024@1Re. QA 25.024110786561720 Gemma 4-26B A4B 1024@100 Re. QA 43.9263242924117 11314 Gemma 4-26B A4B 1024@50Re. QA 43.4197942844036 10262 Gemma 4-26B A4B 1024@25Re. QA 43.31126327642579310 Gemma 4-26B A4B 1024@5Re. QA 42.8381214439556256 Gemma 4-26B A4B 1024@3Re. QA 39.8311110043686011 Gemma 4-26B A4B 1024@2Re. QA 36.4276108542015616 Gemma 4-26B A4B 1024@1Re. QA 30.8241107847665532 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 unitsk, 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. ModelBudgetTaskAcc. Input text Input visual Output textTotal Qwen 3.5-9B1024@200 Re. QA 33.214574244803188 44049 Qwen 3.5-9B1024@100 Re. QA 29.56266165183908 28185 Qwen 3.5-9B1024@10Re. QA 24.2782187060429134 Gemma 4-26B A4B 1024@200 Re. QA 42.814688648445413 88282 Gemma 4-26B A4B 1024@100 Re. QA 41.26370422026561 57069 Gemma 4-26B A4B 1024@10Re. QA 30.880443518650 14367 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 unitsk, 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. Gemma 4 E4B and Qwen3-VL-8B are omitted for the reasons given in Table 5.

Instant research discovery