Spatial perception
Recognize relations, depth, and orientation already present in an image.
A proposed AAAI-27 workshop
Geometric & Spatial Reasoning in Vision-Language Models
Overview
Vision-language models struggle to maintain geometric consistency across fragments, viewpoints, and reasoning steps. This workshop examines these limitations and methods for improving spatial reasoning as task complexity increases.
The program brings together researchers in multimodal learning, embodied AI, and constraint reasoning through invited talks, contributed papers, posters, a competition, and a panel.
Spatial construction
Construct and maintain globally consistent structures under geometric constraints.
Topics of interest
01Spatial perception & grounding+
Spatial relations, depth, size, orientation, perspective-taking, and generalization across viewpoints.
02Construction under constraints+
Combining appearance evidence with geometric constraints in reassembly, multi-step assembly, and grid reasoning.
03Architectures & mechanisms+
Visual tokenization, resolution, positional encodings, spatial memory, combinatorial decoding, and locating where geometric information is lost.
04Neuro-symbolic reasoning+
Pairing VLM proposals with constraint solvers, search, and classical methods for geometric compatibility.
05Training & representations+
Spatial pretext tasks, exactly verifiable construction tasks, and reinforcement learning with verifiable rewards.
06Evaluation methodology+
Uniqueness guarantees, machine-checkable answers, contamination resistance, and metrics that distinguish local from global correctness.
07Embodied AI & applications+
Manipulation, assembly, navigation, and world models—and whether benchmark gains transfer to physical competence.
Invited speakers
Mohit Bansal
University of North Carolina at Chapel Hill
Joydeep Biswas
University of Texas at Austin
Scott Sanner
University of Toronto
Ashok Goel
Georgia Institute of Technology
Call for papers
We invite research and position papers on the topics above.
This workshop is non-archival.
Long papers
Up to 7 pages of main content, with no page limit for references and appendices.
Short & position papers
Up to 4 pages of main content and up to 2 pages of references.
Previously published papers
Recently published work, for presentation only. A separate submission process for these papers will be announced later.
Every accepted paper will be presented as a poster. Selected papers will also receive an oral presentation slot.
Spatial reconstruction competition
Recover the original positions of shuffled, interlocking image pieces using visual evidence and geometric constraints.
2,000Held-out instances
16–256Pieces per instance
4Difficulty levels
Each instance is verified by a constraint solver to have exactly one solution.
Opens 4 months before the workshopCloses 3 weeks before the workshop
Planned competition · Dates, data, baselines, and evaluation code to be announced.
COMPETITION TRACKS
Zero-shot prompting
Use API-only models with zero-shot prompting.
Open-weight models
Fine-tune open-weight models or apply reinforcement learning post-training under a declared compute budget.
Hybrid neuro-symbolic methods
Combine vision-language models with external constraint solvers.
Program
Tentative scheduleWELCOME
Opening remarks & problem framing
INVITED TALK I
Joydeep Biswas
University of Texas at Austin
CONTRIBUTED RESEARCH
Oral presentations I
Four talks · 15 minutes each
POSTER SESSION
Coffee break & poster session I
INVITED TALK II
Scott Sanner
University of Toronto
CHALLENGE
Competition session & results
INVITED TALK III
Ashok Goel
Georgia Institute of Technology
CONTRIBUTED RESEARCH
Oral presentations II
Four talks · 15 minutes each
INVITED TALK IV
Mohit Bansal
University of North Carolina at Chapel Hill
POSTER SESSION
Coffee break & poster session II
PANEL DISCUSSION
Learn to satisfy constraints—or learn when to delegate to a solver?
CLOSING
Open discussion, best paper award & closing
Organizing committee
Shawn Li
University of Southern California
Yue Zhao
University of Southern California
Leman Akoglu
Carnegie Mellon University
Xiyang Hu
Arizona State University
Emilio Ferrara
University of Southern California
SUPPORTING TEAM
Tianyu Shi
McGill University
Ruolin Li
University of Southern California