Research map / Multi-agent systems and games / LLM agents
R^2-Bench: Benchmarking the Robustness of Referring Perception Models under Perturbations
Xiang Li, Kai Qiu, Jinglu Wang, Xiaohao Xu, Kashu Yamazaki, Hao Chen, Rita Singh, Xiaonan Huang, Bhiksha Raj
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 |
|---|---|---|---|
| use polygon-vertex sequential mask prediction (PolyFormer) instead of per-pixel classification | referring image segmentation models · RefCOCO/+/g under perturbation | mIoU and APC: RefCOCO APC -0.12 vs LAVT -0.21 | Improved |
| multi-task joint training (SEEM) | referring image segmentation · RefCOCO | APC: clean mIoU 0.65 but APC -0.12 | Mixed |
| decoupled visual and ID embeddings (DeAOT) instead of pixel/object correspondence (Cutie) | video object segmentation · YTVOS | G APC: G_s APC -2.7 (DeAOT) vs -4.1 (Cutie) | Improved |
| unseen object categories instead of seen | VOS models · YTVOS | performance drop: AOT APC -5.4 seen vs -7.7 unseen | Hurt |
| frame-by-frame feature processing (OnlineRefer) instead of relying on temporal correspondence | referring VOS with image-level backbone · Ref-YTVOS | J APC: -9.4 vs ReferFormer -17.4 | Improved |
| quantization and local-global distillation for acoustic representation (QSD) | audiovisual segmentation · AVS-s4 | F APC: F APC -7.9 | Improved |
| frost perturbation vs color-based (brightness, saturate) perturbations (ablation) | perturbation types · visual perturbations | performance degradation | Hurt |
| acoustic perturbations on AVS (ablation) | AVS models · AVS-s4/ms3 | performance: e.g. AVSegFormer 81.5-81.6 across most types | No clear effect |
| impulse response acoustic perturbation (ablation) | AVS models · AVS-s4 | performance: QSD 79.1 under IR vs about 81.5 for others | Hurt |
| dynamic (frame-varying) instead of static perturbation (ablation) | referring VOS models · Ref-DAVIS | performance: ReferFormer Ref-YTVOS J -13.7 static vs -17.8 dynamic | Hurt |
| increase R2-Agent data verification iterations from 0 to 2 (ablation) | R2-Agent data sample selection · agent data selection user study | ACC: 0.53 -> 0.77 | Improved |
| increase verification iterations from 2 to 3 (ablation) | R2-Agent perturbation selection · agent data selection user study | ACC: 0.86 -> 0.87 | No clear effect |
Closest papers by meaning
- Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks (ICLR 2026 · similarity 0.87)
- VP-Bench: A Comprehensive Benchmark for Visual Prompting in Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
- Unveiling the Tapestry of Consistency in Large Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Understanding Task Transfer in Vision-Language Models (CVPR 2026 · similarity 0.87)
- A Large-Scale Human-Centric Benchmark for Referring Expression Comprehension in the LMM Era (NeurIPS 2024 · similarity 0.86)
- MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs (ICML 2026 · similarity 0.86)
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.
