LVR @ WACV 2027: Latent Visual Reasoning

About

What should a model think with when words are not enough? Should it keep an internal scene, imagine how objects move, revise a spatial hypothesis, or switch among visual, linguistic, symbolic, and tool-based computation?

Current multimodal large language models typically encode an image as context and carry out intermediate reasoning in language. Language is well suited to abstraction and communication, but it often struggles with precise pose, depth, correspondence, occlusion, and temporal or geometric structure. People, by contrast, often simulate how a scene might unfold rather than narrating every detail. When an apple falls from a table, we can imagine its motion, the collision, and where it lands without putting the whole process into words. That kind of spatial and dynamic reasoning may be essential for visual intelligence in video, 3D vision, robotics, medical imaging, and scientific vision.

The Latent Visual Reasoning (LVR) workshop focuses on visual reasoning that does not rely exclusively on text as the intermediate representation. We welcome work on visual tokens, latent scenes, imagined visual states, and other structured representations that keep task-relevant visual information, as well as hybrid systems that combine these with language, symbols, explicit visual operations, or tools. In every case, the intermediate representation should play a testable computational role, not merely coincide with an ordinary hidden activation.

Invited Speakers

Robert Geirhos

Staff Research Scientist

Google DeepMind

Youngeun Kim

Assistant Professor

Korea University

Min Xu

Associate Professor

Carnegie Mellon University

Call for Papers

We invite submissions on latent visual reasoning: work whose intermediate computation is not limited to text, especially when visual structure is kept and updated across reasoning steps. Topics include, but are not limited to:

Representations of evolving visual state: visual tokens, latent scenes, structured memories, mental imagery, and persistent spatial or temporal states.
Learning and inference: mechanisms that route among latent computation, language, symbols, visual operations, and tools.
Tasks and evaluation: images, video, 3D vision, embodied systems, medical imaging, and scientific vision, including supervision, test-time computation, causal interventions, matched-compute comparisons, benchmarks, faithfulness, interpretability, efficiency, safety, and failure analysis.

Tracks and publication

We welcome submissions in the following three categories:

Archival regular papers (8 pages). Accepted papers will appear in the WACV 2027 proceedings, subject to the conference's final formatting requirements.
Non-archival short papers (4 pages). Accepted papers will be presented at the workshop but will not appear in the WACV 2027 proceedings.
Non-archival presentation-only papers. Work already accepted at another venue is welcome for additional visibility and discussion. Authors should identify the original venue and publication status; these submissions will not be republished.

All submissions will be handled through OpenReview. Submission portal: Submit via OpenReview. Formatting: please use the official WACV 2027 LaTeX Author Kit and follow the WACV 2027 Author Guidelines.

Timeline

Event	Date
Workshop website live by	August 25, 2026
Submission deadline	October 10, 2026 (AoE)
Author notification deadline (archival papers)	October 30, 2026
Metadata of accepted papers due to IEEE (archival papers)	November 2, 2026
Camera-ready deadline (archival papers)	November 20, 2026
Workshop date	January 4 or 5, 2027 (TBD)

Tentative Schedule

LVR will be a half-day, in-person workshop at WACV 2027. Times and room: TO BE ADDED.

Session	Details
Opening and research agenda	TO BE ADDED
Invited talks	TO BE ADDED
Contributed orals and spotlights	TO BE ADDED
Posters and demonstrations	TO BE ADDED
Panel: “How Do We Know a Model Is Reasoning Visually?”	TO BE ADDED
Awards and closing remarks	TO BE ADDED