Research map / Label-efficient and robust learning / Test-time adaptation
Learning from Reliable Negatives: Confidence-Anchored Test-Time Adaptation for GUI Grounding
Yizhou Liu, Fei Tang, Yuchen Yan, Zhengxi Lu, Songqin Nong, Tao Jiang, Wenhao Xu, Wenqi Zhang, Weiming Lu, Jun Xiao, Yongliang Shen
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 |
|---|---|---|---|
| select pseudo-label by coordinate-token confidence instead of all-token confidence (ablation) | CAL/CANL · ScreenSpot-V2 pseudo-label (N=12) | pseudo-label accuracy: 78.6 -> 83.9 | Improved |
| use negative-only learning (CANL) instead of CAL on ScreenSpot-Pro (ablation) | Qwen-2.5-VL-7B · ScreenSpot-Pro | accuracy: 32.7 -> 33.8 | Improved |
| use negative-only learning (CANL) instead of CAL on UI-Vision (ablation) | Qwen-2.5-VL-7B · UI-Vision | accuracy: 18.6 -> 20.1 | Improved |
| use negative-only learning (CANL) instead of CAL on ScreenSpot-V2 (ablation) | Qwen-2.5-VL-7B · ScreenSpot-V2 | accuracy: 92.1 -> 92.1 | No clear effect |
| use negative-only learning (CANL) instead of CAL on AndroidWorld (ablation) | Qwen-2.5-VL-3B grounding · AndroidWorld | success rate: 30.2 -> 29.3 | Hurt |
| negative-only vs positive-only learning (CAPL) (ablation) | Qwen-2.5-VL-3B · ScreenSpot-V2 | accuracy: 86.7 -> 88.5 | Improved |
| increase distance threshold tau from 0.05 to 0.1 (CAL) (ablation) | CAL · ScreenSpot-V2 | accuracy: 88.9 -> 86.9 | Hurt |
| increase distance threshold tau from 0.05 to 0.1 (CANL) (ablation) | CANL · ScreenSpot-V2 | accuracy: 88.5 -> 88.5 | No clear effect |
| use binary reward instead of continuous distance reward (ablation) | CAL · ScreenSpot-V2 | accuracy: 85.2 -> 88.9 | Improved |
| use RL (GRPO) instead of SFT on pseudo-labels (ablation) | Qwen-2.5-VL-3B · ScreenSpot-V2 | accuracy: 84.6 -> 88.9 | Improved |
| apply label-free CAL/CANL test-time training to base model | Qwen-2.5-VL-3B · ScreenSpot-Pro | accuracy: 16.1 -> 32.7 | Improved |
| apply CAL as grounding in AndroidWorld agent | Qwen-2.5-VL-3B + GPT-4o planner · AndroidWorld | success rate: 25.0 -> 30.2 | Improved |
Closest papers by meaning
- GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · similarity 0.91)
- Test-Time Reinforcement Learning for GUI Grounding via Region Consistency (AAAI 2026 · similarity 0.90)
- Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding (NeurIPS 2026 · similarity 0.89)
- UI-Ins: Enhancing GUI Grounding with Multi-Perspective Instruction as Reasoning (ICLR 2026 · similarity 0.89)
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Scaling Computer-Use Grounding via User Interface Decomposition and Synthesis (NeurIPS 2025 · 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.
