Research map / Continual learning / Continual fine-tuning of LLMs
CGL: Advancing Continual GUI Learning via Reinforcement Fine-Tuning
Zhenquan Yao, Zitong Huang, Yihan Zeng, Jianhua Han, Hang Xu, Chun-Mei Feng, Jianwei Ma, Wangmeng Zuo
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 GRPO with KL instead of SFT only (ablation) | Qwen2.5-VL-3B · AndroidControl-CL Order 1 | avg step accuracy / FM: 76.90 -> 81.53; FM -5.73 -> -0.62 | Improved |
| add KL constraint to SFT (ablation) | SFT · AndroidControl-CL Order 1 | avg step accuracy: 76.90 -> 80.84 | Improved |
| remove KL from GRPO (ablation) | GRPO · AndroidControl-CL Order 1 | avg step accuracy: 81.53 -> collapse (unavailable) | Hurt |
| add 5% historical replay to SFT (ablation) | SFT · AndroidControl-CL Order 1 | avg step accuracy: 76.90 -> 79.80 | Improved |
| add static full SFT to GRPO (ablation) | GRPO+KL · AndroidControl-CL Order 1 | avg step accuracy: 81.53 -> 81.68 | Improved |
| replace static SFT with error-aware dynamic SFT (ablation) | SFT+GRPO · AndroidControl-CL Order 1 | avg step accuracy: 81.68 -> 81.90 | Improved |
| add entropy-regulated lambda (ablation) | D-SFT+GRPO · AndroidControl-CL Order 1 | avg step accuracy: 81.90 -> 82.23 | Improved |
| add gradient surgery (ablation) | D-SFT+GRPO · AndroidControl-CL Order 1 | avg step accuracy / FM: 81.90 -> 82.10; FM -0.33 -> -0.05 | Improved |
| use dynamic lambda instead of fixed lambda 1 (ablation) | CGL · AndroidControl-CL Order 1 | avg step accuracy: 81.62 -> 82.33 | Improved |
| use dynamic lambda instead of fixed lambda 0.01 (ablation) | CGL · AndroidControl-CL Order 1 | avg step accuracy: 81.90 -> 82.33 | Improved |
| CGL vs GRPO across three task orders | Qwen2.5-VL-3B · AndroidControl-CL Orders 1-3 | avg step accuracy: GRPO 81.44-81.75 -> CGL 82.33-82.61 | Improved |
| CGL vs SFT forgetting on LLaVA-OV-0.5B | LLaVA-OneVision-0.5B · AndroidControl-CL Order 1 | FM: -6.81 -> -0.52 | Improved |
Closest papers by meaning
- Continual GUI Agents (ICML 2026 · similarity 0.91)
- GuirlVG: Incentivize GUI Visual Grounding via Empirical Exploration on Reinforcement Learning (ICLR 2026 · similarity 0.90)
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- GUI-Shift: Enhancing VLM-Based GUI Agents through Self-supervised Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.88)
- GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · 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.
