VAREX-Bench
Models extract structured fields from 1,777 government forms using a supplied JSON schema. Normalized exact match checks field values, while string similarity gives partial credit. Array entries use order-independent matching.
The forms contain synthetic values and typed English text on single pages. Results separate plain text, spatial text, images, and combined inputs. Schema annotation noise limits comparisons near perfect accuracy.
- Made by
- IBM Research
- Released
- March 2026
- Kind
- Documents, charts and text in images
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use, Generated
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision
- Strongest model tested
-
Gemini 2.5 Flash Sep 2025 #112 on the capability index
- Last model added
- December 2025
- Scores refreshed
- November 18th, 2026
In the authors’ words
We introduce VAREX (VARied-schema EXtraction), a benchmark for evaluating multimodal foundation models on structured data extraction from government forms. VAREX employs a Reverse Annotation pipeline that programmatically fills PDF templates with synthetic values, producing deterministic ground truth validated through three-phase quality assurance. The benchmark comprises 1,777 documents with 1,771 unique schemas across three structural categories, each provided in four input modalities: plain text, layout-preserving text (whitespace-aligned to approximate column positions), document image, or both text and image combined. Unlike existing benchmarks that evaluate from a single input representation, VAREX provides four controlled modalities per document, enabling systematic ablation of how input format affects extraction accuracy – a capability absent from prior benchmarks. We evaluate 20 models from frontier proprietary models to small open models, with particular attention to models <=4B parameters suitable for cost-sensitive and latency-constrained deployment. Results reveal that (1) below 4B parameters, structured output compliance – not extraction capability – is a dominant bottleneck; in particular, schema echo (models producing schema-conforming structure instead of extracted values) depresses scores by 45-65 pp (percentage points) in affected models; (2) extraction-specific fine-tuning at 2B yields +81 pp gains, demonstrating that the instruction-following deficit is addressable without scale; (3) layout-preserving text provides the largest accuracy gain (+3-18 pp), exceeding pixel-level visual cues; and (4) the benchmark most effectively discriminates models in the 60-95% accuracy band. Dataset and evaluation code are publicly available.
Gemini 2.5 Pro (snapshot unspecified) tops the leaderboard with 98% Gemini 2.5 Flash leads with 97.3% among models released by December 2025
Best score per entry from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Gemini 2.5 Flash, at 97.3%. This view shows only models we track, each at its best.
The newest model with a score here came out in December 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 98.0 Gemini 2.5 Pro (snapshot unspecified)
- 97.3
Gemini 2.5 Flash
- 96.6 Qwen3-VL 8B
- 95.6
Llama 4 Maverick
- 94.8 GPT-4o (snapshot unspecified)
- 94.8
Ministral 3 14B
- 94.3
Llama 4 Scout
- 94.2 granite-vision-4.1-4b
- 90.8 NuExtract 2.0 2B
- 85.6 InternVL3.5 2B
- 85.5 granite-4.0-3b-vision
- 82.5
Qwen2.5-VL-3B
- 71.0
Gemma 3n E2B
- 67.9 MiniCPM-V4
- 65.3
Gemma 3 4B
- 61.3 h2oVL Miss. 2B
- 34.2 h2oVL Miss. 800M
- 34.2 Qwen3-VL 2B
- 28.2 InternVL3.5 1B
- 9.70
Qwen2-VL-2B-Instruct
- 1 Gemini 2.5 Pro (snapshot unspecified) Image (V); temperature 0 98%
- 2
Gemini 2.5 Flash Sep 2025, Image (V); temperature 0 97.3%
- 3 Qwen3-VL 8B Image (V); temperature 0 96.6%
- 4
Llama 4 Maverick Image (V); temperature 0 95.6%
- 5 GPT-4o (snapshot unspecified) Image (V); temperature 0 94.8%
- 6
Ministral 3 14B Image (V); temperature 0 94.8%
- 7
Llama 4 Scout Image (V); temperature 0 94.3%
- 8 granite-vision-4.1-4b Image (V); temperature 0 94.2%
- 9 NuExtract 2.0 2B Image (V); temperature 0 90.8%
- 10 InternVL3.5 2B Image (V); temperature 0 85.6%
- 11 granite-4.0-3b-vision Image (V); temperature 0 85.5%
- 12
Qwen2.5-VL-3B Image (V); temperature 0 82.5%
- 13
Gemma 3n E2B Image (V); temperature 0 71%
- 14 MiniCPM-V4 Image (V); temperature 0 67.9%
- 15
Gemma 3 4B Image (V); temperature 0 65.3%
- 16 h2oVL Miss. 2B Image (V); temperature 0 61.3%
- 17 h2oVL Miss. 800M Image (V); temperature 0 34.2%
- 18 Qwen3-VL 2B Image (V); temperature 0 34.2%
- 19 InternVL3.5 1B Image (V); temperature 0 28.2%
- 20
Qwen2-VL-2B-Instruct Image (V); temperature 0 9.70%
Source: Official VAREX-Bench leaderboard
- 1
Gemini 2.5 Flash Sep 2025, Image (V); temperature 0 97.3%
- 2
Llama 4 Maverick Image (V); temperature 0 95.6%
- 3
Ministral 3 14B Image (V); temperature 0 94.8%
- 4
Llama 4 Scout Image (V); temperature 0 94.3%
- 5
Qwen2.5-VL-3B Image (V); temperature 0 82.5%
- 6
Gemma 3n E2B Image (V); temperature 0 71%
- 7
Gemma 3 4B Image (V); temperature 0 65.3%
- 8
Qwen2-VL-2B-Instruct Image (V); temperature 0 9.70%
Source: Official VAREX-Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 2.5 Pro (snapshot unspecified) | 98% | Image (V); temperature 0 | Official leaderboard | |
| 2 | | 97.3% | Image (V); temperature 0 | Sep 2025 | Official leaderboard |
| 3 | Qwen3-VL 8B 8B | 96.6% | Image (V); temperature 0 | Official leaderboard | |
| 4 | | 95.6% | Image (V); temperature 0 | Apr 2025 | Official leaderboard |
| 5 | GPT-4o (snapshot unspecified) | 94.8% | Image (V); temperature 0 | Official leaderboard | |
| 6 | | 94.8% | Image (V); temperature 0 | Dec 2025 | Official leaderboard |
| 7 | | 94.3% | Image (V); temperature 0 | Apr 2025 | Official leaderboard |
| 8 | granite-vision-4.1-4b 4B | 94.2% | Image (V); temperature 0 | Official leaderboard | |
| 9 | NuExtract 2.0 2B 2B | 90.8% | Image (V); temperature 0 | Official leaderboard | |
| 10 | InternVL3.5 2B 2B | 85.6% | Image (V); temperature 0 | Official leaderboard | |
| 11 | granite-4.0-3b-vision 3B | 85.5% | Image (V); temperature 0 | Official leaderboard | |
| 12 | | 82.5% | Image (V); temperature 0 | Jan 2025 | Official leaderboard |
| 13 | | 71% | Image (V); temperature 0 | Jun 2025 | Official leaderboard |
| 14 | MiniCPM-V4 4B | 67.9% | Image (V); temperature 0 | Official leaderboard | |
| 15 | | 65.3% | Image (V); temperature 0 | Mar 2025 | Official leaderboard |
| 16 | h2oVL Miss. 2B 2B | 61.3% | Image (V); temperature 0 | Official leaderboard | |
| 17 | h2oVL Miss. 800M 0.8B | 34.2% | Image (V); temperature 0 | Official leaderboard | |
| 18 | Qwen3-VL 2B 2B | 34.2% | Image (V); temperature 0 | Official leaderboard | |
| 19 | InternVL3.5 1B 1B | 28.2% | Image (V); temperature 0 | Official leaderboard | |
| 20 | | 9.70% | Image (V); temperature 0 | Aug 2024 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 97.3% | Image (V); temperature 0 | Sep 2025 | Official leaderboard |
| 2 | | 95.6% | Image (V); temperature 0 | Apr 2025 | Official leaderboard |
| 3 | | 94.8% | Image (V); temperature 0 | Dec 2025 | Official leaderboard |
| 4 | | 94.3% | Image (V); temperature 0 | Apr 2025 | Official leaderboard |
| 5 | | 82.5% | Image (V); temperature 0 | Jan 2025 | Official leaderboard |
| 6 | | 71% | Image (V); temperature 0 | Jun 2025 | Official leaderboard |
| 7 | | 65.3% | Image (V); temperature 0 | Mar 2025 | Official leaderboard |
| 8 | | 9.70% | Image (V); temperature 0 | Aug 2024 | Official leaderboard |
Rankings by part
We read 3 parts from the benchmark’s own leaderboard: Flat, Nested and Table. Each ranks models on one part of the test. Gemini 2.5 Flash leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.
Flat
8 models we track have a score here.
- 1
Gemini 2.5 Flash Sep 2025 96.9%
- 2
Llama 4 Scout ↑2 95.2%
- 3
Llama 4 Maverick ↓1 95.1%
- 4
Ministral 3 14B ↓1 94.8%
- 5
Qwen2.5-VL-3B 79.9%
- 6
Gemma 3n E2B 76.7%
- 7
Gemma 3 4B 71.7%
- 8
Qwen2-VL-2B-Instruct 16.4%
Nested
8 models we track have a score here.
- 1
Gemini 2.5 Flash Sep 2025 97.7%
- 2
Llama 4 Maverick 96.1%
- 3
Ministral 3 14B 95.1%
- 4
Llama 4 Scout 94.3%
- 5
Qwen2.5-VL-3B 83.1%
- 6
Gemma 3n E2B 71%
- 7
Gemma 3 4B 65%
- 8
Qwen2-VL-2B-Instruct 9.00%
Table
8 models we track have a score here.
- 1
Gemini 2.5 Flash Sep 2025 96.5%
- 2
Llama 4 Maverick 94.4%
- 3
Llama 4 Scout ↑1 94.1%
- 4
Ministral 3 14B ↓1 94%
- 5
Qwen2.5-VL-3B 81.9%
- 6
Gemma 3n E2B 68.8%
- 7
Gemma 3 4B 63.6%
- 8
Qwen2-VL-2B-Instruct 8.90%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official VAREX-Bench leaderboard
Llama 4 Maverick on the official leaderboard passed 85% in April 2025
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official VAREX-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the official leaderboard (20 entries).
- Pass Models scored 8 models we track, plus 12 other entries on its leaderboard.
- Note Newest score The newest came 11 months ago, in December 2025, for Ministral 3 14B. That’s expected for a saturated test.
- Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Gemini 2.5 Flash (#112).
- Note Headroom Saturated. The best is 97.3%, past 85% since April 2025.
Leaderboard import 3 pass
- Pass Importer We read the board on udibarzi.github.io: 8 rows for models we track and 12 other entries at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.