Research map / Large language models / Agents, code and math
VIRO: Robust and Efficient Neuro-Symbolic Reasoning with Verification for Referring Expression Comprehension
Hyejin Park, Junhyuk Kwon, Suha Kwak, Jungseul Ok
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 |
|---|---|---|---|
| add operator-based programs (no verification) to detector-only (ablation) | GroundingDINO-T · gRefCOCO no-target + RefCOCO TestA | balanced accuracy: 40.0 -> 56.8 | Improved |
| add logical verification (LV) to operators (ablation) | VIRO · gRefCOCO no-target + RefCOCO TestA | balanced accuracy: 56.8 -> 57.0 | No clear effect |
| add uncertainty verification with fixed threshold (ablation) | VIRO · gRefCOCO no-target + RefCOCO TestA | balanced accuracy: 57.0 -> 58.8 | Improved |
| use adaptive per-label threshold instead of fixed in uncertainty verification (ablation) | VIRO · gRefCOCO no-target + RefCOCO TestA | balanced accuracy: 58.8 -> 61.1 | Improved |
| add verification (LV and UV adaptive) to operators (ablation) | VIRO · RefCOCO testA | TPR (Acc@0.5): 74.6 -> 71.9 | Hurt |
| add uncertainty verification (adaptive) to operators (ablation) | VIRO · gRefCOCO no-target (TestA) | TNR (N-acc): 39.3 -> 50.2 | Improved |
| use CLIP ViT-H/14 instead of ViT-L/14 (ablation) | VIRO · RefCOCO testA | TPR (Acc@0.5): 68.8 -> 71.9 | Improved |
| raise the OVD detection threshold (ablation) | VIRO · RefCOCO testA / gRefCOCO | TPR and TNR | Mixed |
| replace Python-code generation and iterative planners with decoupled symbolic programs and validator | VIRO vs HYDRA · RefCOCO | program failure rate (%): 28.46 -> 0.07 | Improved |
| decouple program generation from execution (amortize one program over N images) | VIRO · 1-query-N-images | total inference time | Improved |
| VIRO vs ViperGPT end-to-end latency (validator loop adds LLM cost) | VIRO · RefCOCO testA | E2E runtime (s/query): 30.19 -> 12.92 | Improved |
| VIRO vs NAVER on RefCOCOg exclusive accuracy | VIRO · RefCOCOg test | Acc@0.5 (exclusive): 68.4 -> 66.6 | Hurt |
Closest papers by meaning
- NePTune: A Neuro-Pythonic Framework for Tunable Compositional Reasoning on Vision-Language (ICLR 2026 · similarity 0.89)
- Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks (ICLR 2026 · similarity 0.88)
- Rex-Thinker: Grounded Object Referring via Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.88)
- GENOME: Generative Neuro-Symbolic Visual Reasoning by Growing and Reusing Modules (ICLR 2024 · similarity 0.87)
- NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models (AAAI 2026 · similarity 0.87)
- SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation (CVPR 2026 · similarity 0.87)
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.
