Efficient Multimodal Question Answering (EMM-QA)

📌 List of Accepted Papers and Shared Challenge Papers

EMM-QA is an ICML 2026 workshop focused on question answering systems that must balance accuracy, efficiency, and adaptability across multiple input modalities. The workshop brings together researchers from academia and industry working on knowledge-intensive multimodal systems that operate under practical resource constraints.

Rather than focusing only on larger models, the workshop emphasizes methods that make multimodal question answering usable in real settings, including retrieval-augmented systems, compact models, efficient inference, and human-in-the-loop evaluation.

This year’s shared challenge explored that idea in a game setting: players and systems answered pyramid-style questions that mixed text with images, then adapted through tossups, bonuses, confidence, and human-computer teaming. We also used adversarial question writing to probe where multimodal models are still brittle, especially on surprising clues that are easy for humans but hard for systems.

For the human-computer matches, the AI teammates shown in the presentation as a grid of model logos were drafted into teams in ranking order: weaker human teams chose first, then the draft continued upward through the standings and reversed if AI teammates were still available. That pairing of roster and draft order is what makes the teammate selection visually meaningful.

QANTA 2026 challenge CSVs are available for Google Sheets: tossup question difficulty and bonus question difficulty.

Scope

The workshop is centered on efficient multimodal question answering. It also welcomes closely related work on multimodal retrieval, reasoning, evaluation, benchmarking, and efficient inference when those contributions are clearly connected to question answering or other knowledge-intensive multimodal tasks.

Like the previous iteration of EfficientQA, which focused on text-only question answering, we will also host a human-computer question answering competition. If you’d like to take part in that part of the competition (it should be fun!), you can either play as a team or write questions.

Workshop Format

The workshop is planned as a one-day event combining:

• Contributed papers
• Poster presentations
• Invited keynotes
• Shared-task highlights
• A live human-computer question answering event

The workshop will also serve as the venue where we announce the winning systems from the QANTA 2026 computer competition.

The presentation also closes with a few reflections: multimodal systems remain strong overall, but they still struggle with surprising visual clues, explanation quality, calibration, and adversarial generation. That makes the competition useful not just as a leaderboard, but as a way to study the jagged frontier between human and machine performance.

Schedule

• Workshop takes place on July 11th. Both poster sessions
• All talk sessions (invited talks, spotlights, challenge talks, awards, etc.) will take place in the ASEM Ballroom 201 at COEX.
• All poster sessions will take place separately in Hall A, poster boards 1612–1617, 1700–1711, outside the workshop room area at COEX. You can review the Hall A plan here.

Legend

• 🟦 Invited Talks
• 🟨 Contributed Paper Spotlights / Best Challenge Team Talk
• 🟧 Poster Sessions

Sponsors/Acknowledgements

• This workshop is partially supported by Horizon EU programme through project ELOQUENCE, grant no. 101135916.