Learn How to Act from Your Own Interactions: On-Policy Self-Distillation for GUI Agents
Yan Zhang, Daiqing Wu, Huawen Shen, Liang Li, Gang Cao, Zhi Gong, Wei Dai, Xiaode Zhang, et al.
cs.AI
Quick verdict
GUI-SD-v2 通过两阶段特权跟进优化和选择性逐步蒸馏，将在线自蒸馏（on-policy self-distillation）扩展到多轮 GUI agent；结果令人鼓舞，但未报告方差或显著性检验。
Summary
该论文通过 GUI-SD-v2 将在线自蒸馏（OPSD）从 GUI 坐标定位（grounding）扩展到多轮 GUI agent，针对先前工作中的两个所谓瓶颈：特权跟随能力弱和特权指导粗糙。两阶段框架在 Qwen3-VL-Instruct-8B 模型上进行训练。阶段 1（特权跟随优化）将特权增强和原始的 rollout 在共享组中进行优化，针对可验证的动作奖励；特权信息是 Kimi-K3 为 MobileForge 轨迹中的错误步骤生成的纠正注释，并通过拒绝采样进行验证。阶段 2（信息特权蒸馏）将 self-teacher 基于每一步的推理特权和记忆特权（来自 Kimi-K3，来源于执行的动作和下一张截图）进行条件化，并通过反向 KL 将这些指导蒸馏到学生模型的决策和记忆 token 中。在 AndroidWorld 上，其报告的 Pass@1 为 67.2% / Pass@3 为 79.3%，略微领先于 GHD-8B (66.5/73.3) 和 GUI-Owl-7B (66.4)；在 MobileWorld (GUI-only 子集) 上，Pass@1 为 25.6% / Pass@3 为 31.6%，而 OpenMobile-8B 为 17.7/24.8 —— 这是更大的差距。在匹配的基准模型中，它优于 SFT、GRPO、朴素 OPSD 和 GUI-SD-v1。主要注意事项：AndroidWorld 的优势很小且未报告方差或显著性检验；评估仅限 Android，且基于单一 8B 模型；整个流水线依赖于 Kimi-K3 特权教师。
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Existing OPSD methods transfer poorly to multi-turn GUI agents because self-teachers follow privileged guidance unreliably and because coarse privileges give no step-level reasoning or memory supervision (Section 1, Figure 1).
- Privilege-following optimization — shared-group rollouts with and without corrective privileges, subset-normalized advantages under clipped GRPO — raises privilege-following accuracy (50.0→62.9, Table 2) and improves final AndroidWorld success after Stage 2 (67.2 vs 60.3-61.2 for SFT/GRPO initialization).
- Step-specific reasoning and memory privileges distilled via reverse KL are complementary: memory guidance mainly lifts information-retrieval tasks (44.4→55.6) while reasoning guidance lifts task-completion tasks (50.6→64.0) on AndroidWorld (Table 3).
- GUI-SD-v2 outperforms the evaluated SOTA methods on both AndroidWorld and MobileWorld in Pass@1 and Pass@3 (Tables 1, Sections 4.2), with the largest external margin on MobileWorld (25.6 vs 17.7 Pass@1).
Strengths
Fairly controlled baseline design
All four training baselines (SFT, GRPO, Naive OPSD, GUI-SD-v1) share the same base model, the same training-task pools, and the same Kimi-K3 supervision source (Section 4.1, Appendix C), so Table 1 deltas are attributable to the training signal rather than data or teacher scale.
Largest gains on the harder benchmark
MobileWorld GUI-only: 25.6 Pass@1 / 31.6 Pass@3 vs OpenMobile-8B's 17.7/24.8 (Table 1) — a roughly 8-point margin that is harder to dismiss than the AndroidWorld results.
Well-factored ablations
Training configuration is documented
Weaknesses
No variance or significance reporting despite repeated evaluation
SOTA table mixes model scales and training recipes
Frontier-teacher confound is acknowledged but not quantified
Missing method details and internal inconsistency in qualitative evidence
Narrow external validity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper extends on-policy self-distillation (OPSD) from GUI grounding (GUI-SD, Zhang et al. 2026b) to multi-turn GUI agents. It attributes the transfer failure to two bottlenecks: weak privilege following (Figure 1b) and coarse privileged guidance (Figure 1c), and proposes GUI-SD-v2, a two-stage framework: (1) Privilege-Following Optimization (PAPO) and (2) Informative Privilege Distillation. Claims are supported on AndroidWorld and MobileWorld (GUI-only subset) at a single model scale (Qwen3-VL-Instruct-8B), with code and data promised.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GUI-SD-v2 and how does it differ from GUI-SD?
GUI-SD-v2 is the successor to GUI-SD, a training method for GUI agents based on on-policy self-distillation, where a model teaches itself under 'privileged' hints it would not have at test time. GUI-SD applied this to GUI grounding (click localization). The v2 extends it to full multi-turn mobile-agent tasks with two stages: privilege-following optimization (learning to actually use corrective hints, via shared-group GRPO rollouts with and without hints) and informative privilege distillation (distilling step-specific reasoning and memory guidance from a privilege-conditioned self-teacher into the policy's own responses).
What results does GUI-SD-v2 report?
Built on Qwen3-VL-Instruct-8B, it reports 67.2% Pass@1 and 79.3% Pass@3 on AndroidWorld, slightly above GHD-8B (66.5/73.3) and GUI-Owl-7B (66.4), and 25.6%/31.6% on MobileWorld's GUI-only subset versus OpenMobile-8B's 17.7%/24.8%. Against identically set up baselines (SFT 50.9, GRPO 52.6, Naive OPSD 50.0, GUI-SD-v1 57.8 on AndroidWorld Pass@1), the two-stage method wins on both metrics and both benchmarks. Note the paper reports no variance estimates or significance tests.
What is 'privilege following' in this paper?
It is the ability of a policy (and later of the self-teacher) to actually act on privileged hint text that is provided only during training. The authors argue prior OPSD methods often ignore such hints, so Stage 1 trains the model with corrective hints on erroneous steps — jointly optimizing rollouts with and without hints against verifiable action rewards — raising privilege-following accuracy from 50.0% to 62.9% (Table 2).
Is code or data available for GUI-SD-v2?
The abstract states 'Code and training data will be publicly released.' The paper does not include repository links in the extracted text, and reproduction additionally requires Android-emulator rollout infrastructure, Kimi-K3 as the privilege teacher, and unstated details (distillation loss weights, analyzer prompts).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Equations were partially garbled by text extraction: the KL direction is rendered as 'D_KL(p_S || p_T)', so the stated reverse-KL reading could not be verified character-by-character.
- The λ_reason/λ_memory loss weights are never stated in the extracted text, and Appendix D contains an unresolved '??' cross-reference to the response template.
- Figures 1-5 could not be analyzed beyond their embedded text; Figure 3(b)'s embedded example text (library-book event) conflicts with the caption's description (phone number).
- The abstract is truncated mid-sentence in the source text; conclusions were drawn from the body and appendices.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.