Challenges

DocInsights 2026 Challenges

Two challenges advancing document intelligence beyond plain text

DocInsights 2026 hosts two complementary challenge tracks: DocSem for document-grounded quantitative reasoning with evidence attribution, and Dr.DocBench for expert-level document parsing across complex visual and structural content.

Competition Season

August 3–October 10, 2026

DocSem concluded on September 10 and its final test leaderboard is now public. Dr.DocBench remains open until October 10 at 12:59 PM UTC. Each official challenge portal is the system of record for submissions and final rules.

USD 5,000+ Total prize pool
2 tracks Distinct challenge tasks
Workshop pathway System papers and presentations

System Paper Submissions

Shared-task paper submissions are open

Participants in DocSem or Dr.DocBench can still submit system papers. Describe the system, data, models, prompts, tools, evaluation choices, and lessons learned from your challenge participation.

Authors may indicate an archival or non-archival preference; the review committee will make the final archival/non-archival decision. Selected contributions will be invited to present at DocInsights 2026.

September 15, 2026 at 11:59 PM UTC Submit on OpenReview

Challenge 1

DocSem

Document-grounded quantitative reasoning with evidence attribution

Concluded · Final results released

Participants receive a PDF document and a paraphrased query. Systems must identify the relevant quantitative passage, derive the requested answer from the supplied document, and return the visible PDF block IDs that support the prediction.

Dataset and guide Final test leaderboard Canonical source
Training data
908 labelled training tasks with PDFs, answers, and evidence block IDs.
Validation data
217 validation tasks with organizer-held labels and a provisional public validation leaderboard. Final rankings use a separate held-out test set.
Test data
1,730 held-out test tasks and PDFs without labels. Test submissions are closed; the final test leaderboard is public and uses each account's best eligible attempt.

Challenge 2

Dr.DocBench

Expert-level parsing of complex, real-world documents

Submissions open

Dr.DocBench challenges systems to recover structured content from complex document pages, including text, tables, formulas, reading order, and specialized notation. Challenge predictions use a single-page unit and produce structurally faithful Markdown.

Enter on EvalAI Challenge website Dataset and guide Benchmark paper
Evaluation
Text Edit Distance, Table TEDS, Formula CDM, and Reading Order are normalized into an Overall score. Teams are ranked by Overall score.
Submission
Upload a validated submission ZIP containing predictions.jsonl or the canonical mds/ tree. Strict checks reject missing, extra, duplicate, or invalid predictions.
Participation
Open worldwide with no team-size limit. Teams may submit up to 3 times per day and select up to 2 submissions for final evaluation.

Competition Timeline

Track milestones

DocSem and Dr.DocBench have separate final submission deadlines. Use the official portal for each track's exact submission status.

August 3–September 10, 2026

DocSem Competition

The competition concluded on September 10. The final test leaderboard was released on September 11, using each Hugging Face account's best eligible attempt.

August 10–October 10, 2026

Dr.DocBench Competition

Submit through EvalAI. Final evaluation runs October 11–23, with winners announced at DocInsights 2026.

August 31, 2026

DocSem Training Data Update

Seven annotation inconsistencies were corrected in the training split following community feedback. The task definition and data format are unchanged.

September 3, 2026

DocSem Validation GT Refresh

Three organizer-only validation ground-truth labels have now been corrected following additional data review. All existing submissions were rescored and the leaderboard updated.

September 5, 2026

DocSem Test Data Release

1,730 held-out test tasks and PDFs were released without labels. The test submission window ran September 5–10 Anywhere on Earth and is now closed.

September 11, 2026

DocSem Final Test Leaderboard Release

The final test leaderboard is public, showing each Hugging Face account's best eligible attempt and Joint Exact Accuracy.

September 15, 2026

Shared-task Paper Submission Deadline

Submit a system paper for DocSem or Dr.DocBench. Authors may indicate an archival or non-archival preference; the review committee will make the final archival/non-archival decision.

From Competition to Workshop

Share systems, findings, and lessons learned

Challenge participants can still submit concise system papers for workshop consideration. Authors may indicate an archival or non-archival preference; the review committee will make the final archival/non-archival decision. Selected contributions will be invited to present their approaches and findings at DocInsights 2026.

Share your challenge work

DocSem has concluded and Dr.DocBench remains open. Participants in either challenge may submit a system paper for workshop consideration.

Document the system

Participants should report models, data, tools, prompts, and evaluation choices needed to understand and reproduce their submission.

Present selected work

Selected participant contributions will have a pathway to share their work with the workshop community.

Prize and rules notice

The combined prize pool exceeds USD 5,000, including up to USD 3,000 for Dr.DocBench. Consult each official challenge portal for track-level eligibility, team rules, ranking procedures, and award conditions.