Groundbench: Multi-Resolution Polygon Grounding Exposes the Geometry Gap in Vision-Language Models

Referring segmentation in overhead imagery is inherently relational: a query may ask for the buildings north of the road or the pond closest to a residential area, so the correct referent can contain one object, several objects, or none. Existing benchmarks mainly score mask overlap, which cannot verify whether a model actually resolved the stated spatial relation. We introduce GeoRefer-Bench, a benchmark for verifiable geospatial referring segmentation. Each query is represented by an executable logical form over a metric scene graph, and predictions are evaluated with Exact Query Success (EQS), which is satisfied only when the returned instance set exactly matches the set denoted by the query. GeoRefer-Bench contains 700 whole 2048x2048 UAV scenes (2.94 Gpx) at 12.5 and 25 cm ground sampling distance, 26,217 instances, 142,796 spatial relations, and 20,916 executable queries spanning five reasoning levels. It further includes three paraphrases per query, 24.0% unanswerable queries, 2,477 counterfactual pairs, and five leakage-controlled evaluation splits. An independent audit re-derives object geometry, mask ownership, relation values, query execution, and split provenance, finding zero issues across all 700 scenes. Relation-blind strategies can retain non-trivial mIoU while achieving at most 22.7 EQS overall, showing that overlap alone does not certify relational grounding. Across fifteen current models, the strongest reaches 74.1 EQS but drops from 98.9 at level 1 to 60.5 at level 5, while ten models score below 5 EQS on two-hop queries. GeoRefer-Bench turns geospatial referring segmentation from mask matching into verifiable reference resolution.

Visual grounding with multimodal large language models is commonly formulated as autoregressive coordinate generation, where a model outputs bounding-box coordinates as text given an image and a referring-expression prompt. While this interface is simple and compatible with instruction following, it introduces a mismatch between training and evaluation: training optimizes token-level likelihood over coordinate strings, whereas grounding quality is measured by geometric overlap. We propose IoU-PD, an IoU-aware privileged distillation method for coordinate-generating multimodal large language models. IoU-PD uses ground-truth boxes not only as coordinate targets, but also as privileged training-time guidance. During training, the student receives the original image and prompt, while a frozen teacher receives a box-marked image and an augmented prompt that indicates the marked region. The student is trained with a supervised fine-tuning anchor and a privileged distillation loss whose token weights reflect both geometric importance and teacher reliability. At inference time, IoU-PD requires no box overlay, privileged hint, teacher branch, or additional prediction module. Experiments on standard referring-expression grounding benchmarks show consistent region-level improvements over strong coordinate-generating baselines, demonstrating that ground-truth boxes can provide useful privileged guidance beyond serving as coordinate labels. Project page: https://xyzzzh.github.io/IoU-PD/

Vision--language models can identify the correct referent while returning an imprecise bounding box. We study whether a frozen direct-answer model can use its own prediction to allocate one additional localized observation without accessing target annotations at inference. Label-free precision refinement (LFPR) routes predicted-small regions to a higher-resolution pass, re-grounds the expression inside a context crop, admits a candidate only under fixed geometric guards, and returns a fixed coordinate-wise midpoint. We report results across three evidence tiers. On 31,921 retrospective Ref-L4 expressions, LFPR raises mAcc$_{0.5:0.95}$ from 72.947% to 76.013% ([email protected] 88.531%$\to$89.725%, [email protected] 55.788%$\to$61.142%). A frozen transfer to 30,969 RefCOCO/RefCOCO+/RefCOCOg expressions improves every dataset at [email protected], mAcc, and mean IoU (pooled mAcc $+0.645$, [email protected] $+0.817$), while [email protected] is unchanged overall: routing alone gains $+1.162$ points there, but crop, guards, and fusion give back $-1.192$, offsetting rather than showing no strict-IoU effect. A prospective, image-disjoint Flickr30K Entities evaluation improves every endpoint (mAcc $+0.973$, [email protected] $+1.022$), more strongly under a single-box variant (mAcc $+2.575$, [email protected] $+3.689$). The same operator applied to two released grounding specialists improves every endpoint ([email protected] $+1.569$/$+6.716$ for EGM-4B/8B) at roughly twice the latency, composing with specialist training rather than replacing it. A genuine unguarded control (guard removed from the same candidates) underperforms the incumbent on every metric, showing the guard is load-bearing. Together, these results show that referent selection and boundary precision are partially separable, with different components moving opposing regions of the IoU curve -- behavior a single threshold cannot reveal.