Research map / Vision-language and multimodal / Multimodal LLMs
DeCoPatch: Revealing Causal Latent Subspaces in Vision-Language Models for GUI Grounding
Yongkang Zhang, Linjia Kang, Zhimin Wang, Duo Wu, Zhi Wang
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 |
|---|---|---|---|
| patch top-k selected late-layer FFN neurons at decode time | Jedi-7B-1080p · OSWorld-G overall | accuracy (%): 53.19 -> 54.26 | Improved |
| patch selected neurons instead of random neurons (ablation) | Qwen2.5-VL-3B · OSWorld-G overall | accuracy (%): 26.95 -> 29.79 | Improved |
| patch top-k neurons | InternVL2.5-7B · ScreenSpot-v2 avg | accuracy (%): 79.32 -> 84.91 | Improved |
| patch top-k neurons | Pixtral-12B · ScreenSpot-v2 avg | accuracy (%): 36.95 -> 46.23 | Improved |
| patch top-k neurons | UI-Venus-Ground-7B · ScreenSpot-v2 avg | accuracy (%): 93.87 -> 94.50 | Improved |
| patch top-k neurons versus overlaying SoM markers | Jedi-7B-1080p · ScreenSpot-Pro avg | accuracy (%): base 39.29; SoM 40.76; ours 40.09 | Mixed |
| patch top-k neurons | Qwen3-VL-8B · ScreenSpot-Pro avg | accuracy (%): 46.50 -> 48.96 | Improved |
| patch top-k neurons | Qwen3-VL-8B · AndroidControl-Curated-Hard SR | success rate (%): 31.57 -> 33.96 | Improved |
| patch top-k neurons versus SoM | Qwen2.5-VL-7B · AndroidControl-Curated-Hard SR | success rate (%): SoM 31.86 vs ours 31.50; base 29.69 | No clear effect |
| replace selected neurons with random neurons (ablation) | DeCoPatch · ScreenSpot-v2 avg | accuracy (%): InternVL2.5-7B 84.91 -> 77.52 | Hurt |
| suppress (divide) the selected neurons instead of boosting them (ablation) | DeCoPatch · OSWorld-G (hyperparameter sweep) | accuracy (%) | Hurt |
| vary k, alpha and number of late layers (ablation) | DeCoPatch · OSWorld-G | accuracy (%) | Improved |
Closest papers by meaning
- Trifuse: Enhancing Attention-Based GUI Grounding via Multimodal Fusion (ICML 2026 · similarity 0.89)
- Learning GUI Grounding with Spatial Reasoning from Visual Feedback (ICML 2026 · similarity 0.88)
- FocusUI: Efficient UI Grounding via Position-Preserving Visual Token Selection (CVPR 2026 · similarity 0.88)
- GUI-Eyes: Tool-Augmented Perception for Visual Grounding in GUI Agents (AAAI 2026 · similarity 0.88)
- GUI-Spotlight: Adaptive Iterative Focus Refinement for Enhanced GUI Visual Grounding (ICML 2026 · similarity 0.88)
- GuirlVG: Incentivize GUI Visual Grounding via Empirical Exploration on Reinforcement Learning (ICLR 2026 · similarity 0.88)
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.
