acceptodds
Withdrawn from ICLR 2027

LiteGUI: Lightweight GUI Agents via Multi-solution Guided Distillation and Dual-level Reinforcement Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present LiteGUI, a new framework for building lightweight GUI agents. GUI interaction poses unique challenges due to the long-horizon nature of complex tasks and the existence of multiple valid interaction paths, which are difficult for lightweight GUI agents to handle effectively. To address these challenges, LiteGUI introduces a two-stage post-training framework operating on logged GUI states. First, we propose Guided Multi-Solution On-Policy Distillation, which provides training-time privileged teacher guidance at each GUI state by selecting the most-matched valid action from human-verified multi-solution action annotations, while leaving the student's on-policy rollout unchanged. Second, we develop Multi-Solution Dual-Level GRPO, which combines action-level supervision with history-conditioned, per-state planning-quality supervision while accounting for multiple valid actions at each logged GUI state. Together, these stages support multi-step GUI tasks without requiring the student to reproduce fixed demonstration trajectories. We further develop a scalable data generation pipeline and a corresponding multi-solution dataset, Lite-Dataset, to support the training and evaluation of GUI agents, addressing a gap in the existing literature. Extensive experiments across multiple GUI benchmarks demonstrate that LiteGUI substantially improves the performance of lightweight GUI agents over existing training paradigms, including SFT, on-policy distillation and GRPO, and achieves competitive or superior performance compared with substantially larger models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI Agent, Distillation, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HpLnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents LiteGUI, a post-training framework for lightweight GUI agents that combines guided on-policy distillation with multi-solution, dual-level GRPO. The method uses human-verified sets of valid actions to provide teacher guidance during distillation and to reduce false-negative rewards during RL, while an additional planning reward supervises long-horizon task progress. The authors also introduce Lite-Dataset, containing multi-step trajectories and multi-solution annotations, and Lite-Bench for evaluation.

Critical strengths

The multi-solution formulation is well motivated for GUI tasks, where several actions can legitimately advance the same task. LiteGUI uses these alternative actions consistently in both the distillation stage and the GRPO reward, reducing the reliance on a single fixed demonstration path.

Critical weaknesses

1.The largest performance gains are observed on Lite-Bench, while the improvements on the external benchmarks are more modest.

2.For ScreenSpot-Pro, which is primarily a single-step grounding benchmark, different coordinates within the same target element are already treated as valid by the grounding objective. It is therefore unclear what additional ambiguity the multi-solution formulation is intended to address on this benchmark.

Questions for authors

Why is a different training dataset used for ScreenSpot-Pro in the SOTA comparison? Could the authors clarify how much of the gain comes from LiteGUI itself versus the choice of training data?

Reviewer_dA9KOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LiteGUI, a two-stage training framework for GUI agents consisting of Guided On-Policy Distillation and Multi-Solution Dual-Level GRPO. Guided-OPD provides the teacher with a privileged reference action selected from human-verified valid action sets via a most-matched mechanism, while MD-GRPO combines a multi-solution action reward with a VLM-judged planning reward to handle multi-path GUI tasks. The paper also introduces Lite-Dataset and Lite-Bench. Experiments achieves competitive or superior performance compared with substantially larger models.

Critical strengths
  1. The paper correctly identifies that GUI tasks admit multiple valid action paths, and single ground-truth supervision causes false negatives. The multi-solution action reward and multi-solution dataset address a genuine gap in the community.

  2. The two-stage paradigm is designed with each component addressing a specific challenge: guided distillation stabilizes teacher supervision under noisy student outputs, while dual-level rewards jointly support action-level and planning-level learning.

  3. The paper releases Lite-Dataset and Lite-Bench, both of which fill gaps in the community and can support future research on multi-solution GUI agent training and evaluation.

Critical weaknesses
  1. The Most-Matched-GT Guided OPD selects a valid action based on student's output, but how does the method ensure the selected action is the correct reference? Consider a case: the student outputs an action with the correct action type but a wrong position (e.g., click at an incorrect coordinate). Suppose the valid action set contains two candidates: (i). the wrong action type but a position close to student's output (ii). the correct action type but a position farther from the student's output. The current matching function selects between them using weights based on Appendix A.2, yet it remains unclear which candidate would be chosen and whether the choice isappropriate. Although Appendix A.2 provides these parameter values, the method's performance may heavily depend on this weight configuration, and the paper provides no sensitivity analysis to address this concern.

  2. Beside, as the paper acknowledges a single GUI state may admit multiple valid solutions, yet the valid action set is only a finite approximation curated by human annotators. There is no guarantee that this set covers different valid actions for a given state. If the student produces a correct action that falls outside the annotated set, the matching function has no valid reference to align with. How can the method ensure that the valid action set is comprehensive to avoid misguiding the teacher when the student's output is valid but unannotated?

  3. The Multi-Solution Dual-Level GRPO computes reward exclusively from per-state feedback: action reward and the planning reward from a frozen Qwen3-VL-32B judge. No terminal task-success participates in reward computation. If only per-step quality is rewarded, how can the method guarantee that optimizing local action and plan quality translates to overall task completion? A step may appear valid in isolation yet lead to a dead end in the trajectory. Moreover, I guess the planning reward mainly depend on juedge model's performance, which also becomes a concern for the whole trajeactory execution.

For the claim lightweight GUI agent:

  1. The training pipeline is not lightweight. LiteGUI requires a 32B teacher model for on-policy distillation, a frozen 32B VLM judge for planning reward, a two-stage training paradigm, and human-verified multi-solution annotations. This infrastructure cost far exceeds standard SFT or vanilla RL, contradicting the lightweight framing of the method.

  2. No lightweight analysis is provided. The paper does not report training cost, inference latency, or the overhead of the teacher and judge calls relative to the student's own computation. Without comparing resource consumption against SFT, vanilla OPD, or vanilla GRPO under identical settings, the claim that LiteGUI is a lightweight training paradigm remains unsubstantiated.

  3. The 30B-A3B model contradicts the lightweight positioning. The paper's motivation emphasizes the need for 2B/3B-scale agents, yet a 30B model is used alongside the 2B to chase SOTA numbers. Its inclusion suggests the method's gains may depend on model scale rather than the proposed training paradigm itself.

For experiment:

  1. Screenspot-Pro is a GUI grounding benchmark, which is not related to long-horizon task, hence the results do not fully demonstrate the effectiveness of this method in solving long-horizon task.

  2. I suggest adding more same-size model baselines and evaluating on long-horizon benchmarks such as WebArena to better validate the method's effectiveness.

Questions for authors

This paper presents solid work: it proposes a training framework and introduces a new benchmark. However, several concerns in the method design remain insufficiently addressed. Furthermore, the lightweight GUI agent positioning does not fit this paper well: the training pipeline requires a 32B teacher and judge, the experimental evaluation includes a 30B model, and no resource analysis is provided to substantiate the lightweight claim.

Reviewer_BsoZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LiteGUI is a two-stage post-training framework for small GUI agents. First, Guided On-Policy Distillation selects the valid annotated action that best matches the student's current output and provides it only to the teacher. Second, Multi-Solution Dual-Level GRPO rewards both matching any valid action and producing a reasonable subtask plan. The paper also introduces Lite-Dataset and Lite-Bench. Experiments on ScreenSpot-Pro, OS-World, and Lite-Bench show clear gains over SFT, standard OPD, and standard GRPO.

Critical strengths
  • Using multiple valid actions is well motivated for GUI tasks. The most-matched teacher guidance and max-over-solutions reward reduce false penalties for alternative correct actions.

  • The controlled comparison uses the same backbone and data, and the ablations cover the main components. The paper also reports five evaluation runs and useful training-cost and judge-calibration details.

Critical weaknesses
  • MD-GRPO scores candidate actions at saved GUI states without executing them. It therefore directly trains per-state action and plan prediction, rather than online exploration or recovery from action outcomes. More evidence is needed to support the long-horizon exploration and error-recovery claims.

  • The method is described as SFT-free, but both training stages rely on 11K human-verified multi-solution annotations. The paper compares against standard SFT, but not a multi-solution SFT baseline using the same annotations. It is therefore unclear whether the gains come from the proposed training method or the richer supervision.

  • The broader SOTA comparisons are not fully controlled because the models use different training data and evaluation settings. These issues do not weaken the same-data ablations, but the generalization claims should be stated more carefully.

Questions for authors
  • What evidence shows that logged-state MD-GRPO improves online exploration and error recovery?
  • Can the authors compare against a multi-solution SFT baseline trained with the same human-verified annotations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.