Complete structured cross-image relation and trajectory annotation

Complete the structured task-level relation annotations and trajectory-stage annotations for SWE-PolyVision, including participating image identifiers, cross-image constraints, earliest supporting trajectory spans, uncertainty labels, and adjudicated failure categories.

Background

SWE-PolyVision is screened for cross-image relevance, but the paper distinguishes this screening from a more detailed annotation of which images participate in which relation and how agents move through the proposed Observe–Align–Integrate–Abduce–Verify–Repair stages. The structured annotation is intended to support frequency estimates, trajectory audits, and failure analysis, but it had not been completed for the current release.

References

Human construction review screened for at least one clear cross-image relation or constraint, but the structured task-by-task labels assigning images and constraints to these families are not yet complete; reported family descriptions are therefore a taxonomy, not frequency estimates.

— SWE-PolyVision: Benchmarking Cross-Image Abductive Reasoning for Repository-Level Software Engineering  (2609.29754 - Wu et al., 24 Sep 2026) in Section 3, subsection “Cross-Image Reasoning Target”; Appendix, Section “Cross-Image Relation and Trajectory Categories”