Research map / Large language models / Reasoning and chain of thought

Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GUI-SD, an on-policy self-distillation framework for GUI grounding that uses a visually enriched privileged context (red bounding box plus Gaussian soft mask) and entropy-guided token weighting, improving Qwen3-VL-8B over GRPO variants with about 4x faster training per epoch.

Problem

GRPO-based RL for GUI grounding needs expensive multiple rollouts and gives sparse rewards on hard samples; naive on-policy self-distillation with textual privileged coordinates collapses the teacher distribution to near one-hot targets and distributes optimization indiscriminately across coordinate

Approach

Replace textual ground-truth coordinate privilege with a visual privileged context (drawn bounding box, hint, and Gaussian soft-mask zoom) so the teacher keeps soft distributions, and reweight the reverse-KL loss per token by digit positional significance and teacher-entropy confidence.

Key results

• GUI-SD reaches 60.7% on ScreenSpot-Pro vs 57.4% for GRPO-Gaussian and 55.6% for naive OPSD on Qwen3-VL-8B.
• Average across six benchmarks rises from 63.5 (base) to 68.4 for GUI-SD, versus 65.9 for the best GRPO variant.
• Training takes 4.2 hours per epoch versus about 16.8 for GRPO variants on 8xA800, roughly 4x faster.

Surprising findings

• Naive textual-coordinate OPSD improves over SFT by only 1.5% on ScreenSpot-Pro, with teacher entropy collapsing to 0.17.

Limitations

• Only evaluated on Qwen3-VL models; scaling to other families not explored.
• Focuses on single-step grounding rather than long-horizon GUI agent tasks.
• Training data is about 7K filtered and rewritten samples from ScaleCUA.

Experimental setup

Datasets: ScaleCUA grounding subset

Benchmarks: ScreenSpot-v2 ScreenSpot-Pro UI-Vision MMBench-GUI L2 OSWorld-G OSWorld-G-Refine

Metrics: accuracy training time per epoch

Baselines: GRPO-Binary GRPO-Distance GRPO-Gaussian Naive OPSD UI-TARS-7B GTA1-7B GUI-Actor-7B JEDI-7B GUI-G^2-7B Propose-then-Critic-8B ZwZ-8B HyperClick-7B

Builds on: on-policy distillation SDPO GRPO GUI-G^2 SE-GUI

Closest papers by meaning

• SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.92)
• GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · similarity 0.92)
• UI-Ins: Enhancing GUI Grounding with Multi-Perspective Instruction as Reasoning (ICLR 2026 · similarity 0.92)
• Test-Time Reinforcement Learning for GUI Grounding via Region Consistency (AAAI 2026 · similarity 0.91)
• GuirlVG: Incentivize GUI Visual Grounding via Empirical Exploration on Reinforcement Learning (ICLR 2026 · similarity 0.91)
• Learning GUI Grounding with Spatial Reasoning from Visual Feedback (ICML 2026 · similarity 0.90)