Thinking with Anchors

arXiv:2608.04424August 2026

Thinking
with Anchors

Grounded and Efficient Document Reasoning

We turn document regions into reusable visual evidence—connecting geometry, semantics, and reasoning so models can show not only what they know, but where that knowledge comes from.

Paper ADOPD 2026 code DocCount code Models & Datasets
Original illustrated document about British school uniforms The same document annotated with numbered visual regions and dashed OCR boxes

ADOPD2026

120Kdocument pages
1.22Mhuman masks
1.24MOCR groups
12visual classes + OCR boxes
9released checkpoints

DocCountBench

442benchmark samples
4counting categories
2,555counted instances

01 · ADOPD 2026

A region becomes useful when it can act as evidence.

Traditional document understanding often stops at locating page elements. Thinking with Anchors treats every meaningful region as a grounded unit that a reasoning system can reference, compose, and verify.

01Geometrybox · polygon
02Modalitytext · visual entity
03Semantic roletitle · chart · logo · …
anchorVerifiable evidencefor grounded reasoning
12 semantic labels Word size partially reflects mask count · Color identifies class · 1,121,285 target masks in all 12 classes
Word cloud of the 12 ADOPD semantic classes, with font size partially reflecting mask count and a distinct color for each class
01

Doc2Box

Recover human-aligned OCR text blocks.

69.1 AP YOLOv12-X
02

Doc2Mask

Segment visual entities with native polygons in a tag-agnostic setting.

69.5 AP RF-DETR-Seg-2XL
03

Doc2Tag

Infer document-specific semantic roles.

12 labels

Best fine-tuned result shown for each task

02 · DocCount benchmark

From visual anchors to grounded semantic counting.

DocCount tests whether models can apply explicit semantic definitions, ground every qualifying region, and produce a verifiable count on dense document pages.

Data construction pipeline
ADOPD 2026 pipeline showing tagged visual polygons and OCR blocks, enriched page information, and anchor-grounded chain-of-thought reasoning

02 · DocCount benchmark

Can a model count what a definition means?

DocCount tests semantic counting in real document images. Every class is defined by inclusion, exclusion, and counting rules. Every answer is backed by polygon-level ground truth.

Document page with seven numbered logo anchors Original travel advertisement without anchor overlays Target class · Brand Logo
ANCHOR-CoT / verified trace
01

Interpret the definition
Count distinct standalone brand marks. Exclude decorative typography and product imagery.

02

Ground the candidates
Anchors [[1]] through [[7]] each identify a qualifying logo region.

03

Verify and answer
No anchor is duplicated or excluded by the grouping rule.

FINAL ANSWER: 7
442questions & original images
4semantically defined classes
2,555counted instances
72.85%best exact-count accuracyKimi K2.5

DocCount results

State-of-the-art VLMs still miss
roughly one in four pages.

Results use exact integer accuracy over all 442 questions.

Reproduce the evaluation
Kimi K2.5
72.85
Qwen3.6 35B A3B
72.17
Qwen3.5 397B A17B
70.81
GPT-5.5
68.78
Claude Sonnet 4.5
68.10
GPT-5.2
66.29
GLM-4.6V
64.48
Gemma 4 31B IT
60.86
Exact-count accuracy (%) · best complete run per model