Multimodal Retrieval
Scaling search across images, video, charts, and UI with hybrid dense/lexical methods.
Call for Papers
At GRAIL-V we bring together CV, IR, NLP, and HCI communities to advance unified methods and evaluation practices for visual–text search, multimodal tool use, and calibrated decision-making. This is the CVPR 2026 workshop CFP for multimodal generative/understanding models and agents that plan, retrieve, reason and verify with an evidence-centric systems.
We invite papers and demos on grounded vision-language systems that plan, retrieve, reason, and verify with evidence. We welcome submissions that bridge multimodal perception and actionable decision-making. Please follow the CVPR 2026 author guidelines.
Agentic Planning
Understanding tool routing, memory tracking, and safe multi-step workflows.
Generative Handling
Incorporating image/video/text understanding, generation and editing tools into agentic systems.
Grounding & Evidence
Citation provenance, evidence overlays, and audit-ready faithfulness.
Deployment Excellence
Benchmarking reproducibility, latency, and cost efficiency.
Submission details
We invite archival papers and demos on grounded multimodal retrieval, reranking, and verification for agentic vision systems. We especially welcome work that reports grounded evidence (region/page/moment), calibration or abstention behavior, and real deployment constraints (latency, memory, cost).
- Format: CVPR template, up to 8 pages (references excluded).
- Review: double-blind on OpenReview with 3 reviews + AC oversight.
- Evaluation: grounded provenance, calibration/abstention, and efficiency metrics.
- Artifacts: code/data/models encouraged with licenses and provenance.
What you can submit
Archival paper
Up to 8 pages (references excluded), CVPR template, double-blind review, appears in workshop proceedings.
Archival short paper
Early or in-progress ideas that will appear in the CVPR workshop proceedings.
Demo / system paper
Working systems with grounded evidence, evaluation, and reproducible artifacts.
Benchmark / dataset
New tasks, datasets, or evaluation harnesses for grounded retrieval and verification.
How to submit
- Follow CVPR rules Use the CVPR 2026 author guidelines (template, anonymization, ethics).
- Show grounding & efficiency Include region/page/moment evidence, calibration/abstention, and real deployment constraints.
- Include artifacts (encouraged) Link code/data/models with licenses and provenance.
- Submit on OpenReview Use the workshop OpenReview forum.
- Present at the workshop Camera-ready after reviews; at least one author presents in person per CVPR policy.
Review criteria
Technical merit, grounded evaluation, efficiency, reproducibility, and broader safety considerations.
Presentation format
Accepted papers appear as posters or short orals. At least one author presents in person.
Conflict of interest
Same-employer, advisor-advisee, recent co-author, or close personal conflicts.
Important dates
Timeline (Anywhere on Earth)
Mar 5, 2026
CVPR 2026 workshop submission deadline
OpenReview submission closes at 23:59 AoE.
Mar 18, 2026
Notification to authors
Decisions released via OpenReview.
Apr 5, 2026
Camera-ready due
Final versions for CVPR workshop proceedings.
Jun 3-4, 2026
Workshop day in Denver, USA
Full-day program with keynotes, panels, and posters.
Topics of interest
Research that advances grounded, efficient, and safe vision-language systems
Heterogeneous and multimodal retrieval
- Unified visual and text search across images, video, charts, documents, and UI.
- Hybrid dense and lexical retrieval, multilingual search, and long-context documents.
- Region, page, and moment retrieval at fine granularity.
- Retrieval over structured and unstructured sources, including tables and knowledge graphs.
Multimodal reranking
- Cross-encoders and late-interaction variants for grounded selection.
- Throughput and latency trade-offs with distillation, caching, and streaming.
- Calibration and abstention at top-k for agent handoffs.
- Evidence-aware reranking with citations and provenance.
Agentic planning and tool use
- Query reformulation, routing, and tool composition across modalities.
- OCR-free parsing, layout and chart tools, SQL and code tools.
- Safety-conscious tool use with guardrails and recovery.
- Memory and state tracking for multi-step agentic workflows.
Grounding, provenance, and reliability
- Evidence overlays, citation fidelity, and structured alignment.
- Verification protocols that resist leakage and prompt injection.
- Robustness to UI drift, style shifts, and format changes.
- Faithfulness auditing and evidence-backed explanations.
Data enrichment and evaluation
- Layout and structure extraction, temporal alignment.
- Reproducible harnesses, prompts, traces, and leaderboards.
- Shared reporting practices for cross-domain benchmarks.
- Efficiency metrics: latency, memory, energy, and cost.
FAQ
What kinds of papers fit GRAIL-V?
Work on grounded multimodal retrieval, reranking, and verification for agentic vision-language systems: region/page/moment evidence, hybrid structured+unstructured retrieval, tool use, calibration/abstention, robustness, and deployment efficiency.
What should I include to make my submission strong?
Clear evidence grounding (citations or region/page/moment provenance), realistic efficiency reporting (latency/memory/cost), and reproducible artifacts with licenses when possible.
Is GRAIL-V archival?
Yes. Accepted papers appear in the CVPR 2026 Workshop Proceedings (archival).
When is the CVPR 2026 workshop submission deadline?
See the Important Dates section on this page for the exact deadline and timeline updates.
Can I submit non-archival or demo-only work?
Yes. We welcome demos and industry showcases. Indicate this in your submission.
Are remote presentations allowed?
CVPR policy prefers in-person presentations. Remote exceptions require documentation.
Do you require code or data release?
Strongly encouraged. Please include artifact links and licenses when possible.
Stay connected
General inquiries
Reach out to the organizers with questions about submissions, sponsorship, or program.