Research map / Vision-language and multimodal / Multimodal LLMs
MC-Bench: A Benchmark for Multi-Context Visual Grounding in the Era of MLLMs
Yunqiu Xu, Linchao Zhu, Yi Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| feed image sequence instead of merged image | Qwen2-VL-72B · MC-Bench | Acc: 51.4 -> 71.4 | Improved |
| feed image sequence instead of merged image | Qwen2-VL-72B · MC-Bench | AP50: 26.0 -> 30.7 | Improved |
| feed image sequence instead of merged image | Qwen2-VL-7B · MC-Bench | Acc: 51.4 -> 54.9 | Improved |
| feed image sequence instead of merged image | Qwen-VL-Chat 7B · MC-Bench | Acc: 44.9 -> 33.4 | Hurt |
| feed image sequence instead of merged image | InternVL2.5-8B · MC-Bench | Acc: 51.4 -> 36.4 | Hurt |
| scale LLM from 7B to 72B (sequence input) | Qwen2-VL · MC-Bench | Acc: 54.9 -> 71.4 | Improved |
| multi-context instruction tuning | Qwen2-VL-7B · MC-Bench | Acc: 54.9 -> 57.7 | Improved |
| multi-context instruction tuning | Qwen2-VL-7B · MC-Bench | AP50: 19.1 -> 22.6 | Improved |
| use GPT-4o to generate referring phrases and G-DINO to localize instead of GPT-4o end-to-end | GPT-4o · MC-Bench | AP50: 2.8 -> 36.2 | Improved |
| single-image grounding specialist training instead of generalist | specialist MLLMs · MC-Bench | Acc | Hurt |
| add synthetic instance-level data to instruction tuning data (ablation) | Finetuned Baseline · MC-Bench | Acc and AP50 | Improved |
| add only collected image-level multi-context data (ablation) | Qwen2-VL-7B · MC-Bench | Acc and AP50 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Ferret — Ferret: Refer and Ground Anything Anywhere at Any Granularity (ICLR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.94)
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.93)
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.93)
- PostAlign: Multimodal Grounding as a Corrective Lens for MLLMs (ICLR 2026 · similarity 0.92)
- Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.91)
- Grounding Multimodal Large Language Models to the World (ICLR 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
