[Submitted on 14 Oct 2026 (v1), last revised 15 Oct 2026 (this version, v2)]
Abstract:A spatial answer depends on the reference frame specified in language and its supporting visual evidence. Item accuracy alone obscures whether correctness persists when either changes. We introduce Frame Evidence, a benchmark of 320 matched triplets and 960 Chinese multiple-choice questions from captured and rendered scenes. One family compares three grounded reference frames over fixed images. The other pairs image deletion annotated as insufficient with answer-preserving addition. Joint metrics require correct answers across all three frames and paired evidence variants. Across 20 configurations on 288 held-out questions, GPT three-shot reaches 50.00% item accuracy, but the best three-frame group accuracy is 5/48 (10.42%), reached by GPT two-shot. Few-shot prompting and task-specific adaptation improve item accuracy without consistently improving joint correctness. Input controls show where images support matched success, while repeated executions reveal answer instability beneath similar scores.
Submission history
From: Chunxi Luo [view email]
[v1] Wed, 14 Oct 2026 04:22:42 UTC (4,592 KB)
[v2] Thu, 15 Oct 2026 03:11:14 UTC (4,592 KB)
