- Rating
- 1(−1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper introduces MintAct, a unified family of vision-based agents available in 2B, 4B, and 8B parameter sizes. MintAct can directly perceive screen content, execute multi-step operations across mobile devices, computers, and web environments, and utilize external tools. Through supervised fine-tuning, distillation from expert models, and reinforcement learning, MintAct achieves or surpasses the performance of specialized models of similar scale; for instance, the 8B version scored 48.9 on the OSWorld-Verified benchmark. The paper demonstrates that a single model can handle tasks across diverse digital environments, eliminating the need for platform-specific training. Nevertheless, there remains room for improvement in areas such as comprehensive cross-environment joint training, context management for long-horizon tasks, and the seamless integration of interface interactions with tool usage.
Critical strengths
- Agentic reinforcement learning for heterogeneous, multi-platform devices is an intriguing area. Implementing this and achieving a robust deployment requires a significant amount of effort.
Critical weaknesses
- This paper reads more like a project report than a formal research paper. It mainly describes a common pipeline for training a GUI agent model and presents the resulting model. As a result, it is difficult to identify clear novel insights or meaningful contributions to either academia or industry. Overall, it feels less like a research contribution and more like a summary of completed work, reflecting on a project after it has already been carried out.
- Regarding the main contribution claimed in this paper, the base model used to fine-tune MintAct, namely Qwen-3-8B, appears somewhat outdated. Since the newer version, Qwen-3.5-9B, was released in March, it would have been a more suitable choice for a submission to ICLR, given the September deadline. This choice of base model may also make it difficult for the experimental results to remain competitive with those of recent models at a similar scale. If the authors intend to emphasize a methodological contribution, then using Qwen-3-8B may still be acceptable. However, in that case, the paper should provide more substantial insights and analysis of the proposed methodology or system design.
Questions for authors
- In the agentic RL setting, which VLM was used as the judge? In addition, where did the tasks used for training during the agentic RL stage come from? How did the authors verify that these tasks were actually executable in practice?
- To be honest, when I think about agentic reinforcement learning for heterogeneous, multi-platform devices, two questions particularly interest me. First, does joint training across multiple platforms provide any advantages in terms of efficiency, performance, stability, or other aspects, compared with training separate models and then combining them through model merging or other integration methods? Second, in a heterogeneous setting, could differences in execution speed across tasks become more severe during the rollout stage? For example, might tasks in certain domains run more slowly because of their inherent complexity or because of container overhead? If so, could the rollout process become limited by these slower tasks? Moreover, if time limits or early truncation are used, how would that affect performance in those domains? These seem like important trade-offs. How did the authors address them?
3 comments
[W1] Research findings
Thank you for this candid feedback. We respectfully note that MintAct is driven by a research question: whether diverse visual agentic capabilities can be unified within a single compact model without sacrificing performance in any domain, and what makes such unification work. Our experiments address this with several findings, which are reported in Section 4.3 and Appendix A.2 but could be highlighted more clearly. Representative examples include:
- Grounding and multistep SFT are complementary. Either stage alone fails on the other capability: Stage 1 alone reaches only 9.4 on OSWorld-Verified, while Stage 2 alone drops UI-Vision to 25.0. Combining them retains both (Table 3).
- Joint RL transfers across related domains. Joint RL on mobile and desktop improves web navigation without web data (Weblica from 57.5 to 60.8, Table 5).
- Synthetic environments help when real environments are unavailable. RL purely in our synthetic environments, without using either real emulator, improves Qwen3-VL-8B from 47.6 to 60.3 on AndroidWorld and from 33.9 to 38.6 on OSWorld-Verified. When real environments are available, RL on them remains more effective (AndroidWorld 66.7 vs. 59.8, Table B).
We will revise the Introduction to summarize these findings explicitly.
[W2] Results on Qwen-3.5
Thank you for this suggestion. We address it in two ways. MintAct already outperforms the newer Qwen3.5 models at comparable sizes. Table 1 includes Qwen3.5 as baselines. Starting from the older Qwen3-VL base, MintAct-8B outperforms Qwen3.5-9B on AndroidWorld (67.0 vs. 57.8) and OSWorld-Verified (48.9 vs. 41.8), and MintAct-4B likewise outperforms Qwen3.5-4B (62.6 vs. 58.6 and 44.8 vs. 35.6). Our recipe transfers to the newer base model. To verify that our pipeline is not tied to a specific base model, we applied our two SFT stages to Qwen3.5-9B with the same data and hyperparameters (Table R1). Our SFT improves Qwen3.5-9B on all seven benchmarks, with large gains in grounding (UI-Vision +37.1, OSWorld-G +14.7) and consistent gains on mobile and desktop navigation (AndroidWorld +4.6, OSWorld-Verified +1.6). Due to the limited compute during the rebuttal period, we have not yet run the RFT and RL stages on Qwen3.5-9B, and we will include these results in the final version.
| Model | MMBench-GUI | ScreenSpot-v2 | UI-Vision | OSWorld-G | AndroidControl | AndroidWorld | OSWorld-Verified |
|---|---|---|---|---|---|---|---|
| Qwen3.5-9B | 84.7 | 92.8 | 36.9 | 55.0 | 67.6 | 57.8 | 41.8 |
| Qwen3.5-9B + MintAct SFT | 87.8 | 95.2 | 74.0 | 69.7 | 72.1 | 62.4 | 43.4 |
| Δ | +3.1 | +2.4 | +37.1 | +14.7 | +4.5 | +4.6 | +1.6 |
[Q1] RL task source and executability
Thank you for these questions. We build a dedicated data synthesis pipeline for each domain, detailed in Appendix A.1. Taking desktop (OSWorld) as an example (Appendix A.1.1): starting from a small set of human designed seed tasks, a specialized teacher model (EvoCUA-32B) rolls out the tasks (multiple times) in the real OSWorld environment, and our internal VLM proposes new tasks grounded in what these rollouts reveal about the environment. The new tasks are rolled out again in the next round, and we iterate until the tasks cover the environment. Finally, an internal VLM judge evaluates the rollouts and filters out tasks whose rollouts fail, so every remaining task has been successfully executed in the real environment. To select RL tasks of appropriate difficulty from this pool, we roll out our SFT policy eight times per task and keep tasks with both successful and failed outcomes, which provide the most informative learning signal for group based RL.
[Q2 (part 1)] Joint training vs. separate training and merging
Thank you for this insightful question. Among the many ways to combine separately trained models, we adopt the most straightforward one: we train an RL specialist for each domain and distill them back into a single model at the data level through RFT (Section 3.2.3), which clearly improves over the SFT model at all scales (Table 4).
Joint RL further improves this merged model, raising OSWorld-Verified from 43.1 to 48.9 and Online-Mind2Web from 36.3 to 39.1 at 8B, with the same trend at 4B and 2B (Table 4). The two approaches are complementary: merging alone retains only part of the desktop specialist's gain (43.1 vs. 49.2), and joint RL nearly closes this gap (48.9).
Joint training also brings transfer across domains. Starting from the same SFT model, joint RL on mobile and desktop exceeds, on average, every RL specialist trained on a single domain (navigation average 55.9 vs. 54.1 and 54.3). On Online-Mind2Web, it improves to 32.8, while both single domain specialists fall below the SFT model (28.4 and 27.7 vs. 30.0) (Table 5).
[Q2 (part 2)] Heterogeneous rollout speed across domains
Thank you for this insightful question. We agree that task complexity and environment overhead lead to different rollout speeds across domains. Filtering out groups whose rollouts are uniformly successful or uniformly failed widens this gap further, since domains also differ in how often they produce informative groups.
We address this with three mechanisms. First, rollouts are fully asynchronous, so slow trajectories do not block other tasks from progressing. Second, the trainer collects groups under a quota for each domain, and a backpressure mechanism pauses new rollouts from any domain that leads the slowest domain by at least one training batch, so fast domains cannot dominate the training distribution. Third, if a slow domain has not filled its quota after a bounded waiting time, the trainer relaxes the quota and fills the remaining slots with groups from other domains, so a slow domain never stalls training. The trade off is that the slow domain contributes fewer groups to that batch, which could cause domain imbalance if it happened frequently. In our experiments, however, we found that this fallback is rarely triggered.
Since the deadline for rebuttal is approaching, the author has still not submitted a rebuttal. Meanwhile, the other two highly-rated reviews have almost identical weaknesses, suggesting they were AI-generated. Based on the above points and my original review, I'm lowering the score to one.