M$^2$-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining

AI summary

Overview

Research area: Artificial intelligence / GUI (graphical user interface) agents — specifically automated data mining for training mobile GUI agents, using Monte Carlo Tree Search (MCTS) and multi-agent LLM cooperation.

Technical level: Advanced. The paper assumes familiarity with MCTS (selection, expansion, simulation, backpropagation), multimodal large language models (MLLMs), Q-value/UCT-style node scoring, and reinforcement-learning-style data pipelines.

Scope: The paper proposes and evaluates M²-Miner, an automated framework that mines intent–trajectory training data for mobile GUI agents, and shows that agents fine-tuned on the mined M²-Miner-Agent dataset reach state-of-the-art results on several mobile GUI benchmarks.

What This Paper Is About

Training capable mobile GUI agents requires large amounts of high-quality "intent–trajectory" data — a natural-language task description paired with the sequence of screen actions that completes it. Manual annotation is slow and expensive, while existing automated mining methods produce redundant, low-quality, and low-diversity data. M²-Miner's goal is to mine such data automatically, cheaply, and at higher quality and diversity, by embedding a collaborative multi-agent framework and an intent-recycling step inside a tree-search-based mining loop.

Key Contributions

1. M²-Miner, the first automated mobile GUI agent data-mining framework based on MCTS. It introduces a tree-structured representation of the mining process and embeds a collaborative multi-agent framework — InferAgent, OrchestraAgent, and JudgeAgent — for exploration guidance (expansion), acceleration (action merging and ranking), and process-based evaluation (simulation reward).
2. A novel intent recycling strategy. After a tree is mined, root-to-node paths other than the original successful trajectory are filtered, scored, and paired with newly generated intents, extracting extra intent–trajectory pairs without re-mining and increasing intent diversity.
3. A progressive model-in-the-loop training strategy. The InferAgent and JudgeAgent are iteratively retrained across three stages — basic intents (Stage 1), complex intents (Stage 2), and recycled intents (Stage 3) — so that agent capability and mined data complexity grow together, improving mining success rate and generalization to unseen environments.
4. The M²-Miner-Agent dataset and an extensive empirical study. The released dataset contains 20k images and 2,565 trajectories with an average trajectory length of 7.8, and models trained on it (M²-Miner-3B and M²-Miner-7B) achieve SOTA results across benchmarks, validated by ablations on the multi-agent framework, the model-in-the-loop strategy, data structure, and training data composition.

Main Findings

• State-of-the-art GUI agent results. M²-Miner-7B reaches 97.5% TP and 93.5% SR on AC-Low; on AC-High it ranks second in TP (81.8%) but achieves the best SR (72.9%); on AITZ it reaches 81.3% TP and 69.4% SR; on GUI-Odyssey it reaches 90.5% TP and 79.3% SR; and on CAGUI it reaches 88.8% TP and 70.2% SR. M²-Miner-3B reaches 97.2% TP / 93.2% SR on AC-Low, 81.3% / 71.2% on AC-High, 78.6% / 66.6% on AITZ, 88.2% / 77.1% on GUI-Odyssey, and 88.5% / 67.3% on CAGUI.
• Strong generalization to unseen domains. On CAGUI (whose training set is unreleased), training Qwen2.5-VL-7B with the mined data raises SR from 55.2% to 70.2%.
• Better than other auto-mined-data baselines. M²-Miner-7B consistently outperforms OS-Genesis-7B (e.g., 20.0% TP / 8.5% SR on AITZ, 11.7% / 3.6% on GUI-Odyssey) and GUI-Owl-7B (78.9% / 65.1% on AITZ, 83.4% / 60.7% on GUI-Odyssey, 80.0% / 59.2% on CAGUI).
• Comparable to a heavily human-annotated model. UI-TARS-7B (trained on large-scale private human-annotated data) reaches 98.0% TP / 90.8% SR on AC-Low and 88.6% / 70.0% on CAGUI; the paper reports that M²-Miner-7B achieves comparable TP and consistently surpasses UI-TARS-7B in SR across almost all benchmarks.
• Exponential mining-speed improvement. Compared with vanilla MCTS using only the InferAgent, the multi-agent framework yields exponentially growing efficiency gains with task complexity, reaching a 64× speedup at task length 9. The paper attributes this to the OrchestraAgent merging redundant actions and prioritizing correct ones, and the JudgeAgent replacing costly rollouts with direct process-based evaluation.
• Each agent matters. Ablations show that removing the OrchestraAgent or the JudgeAgent reduces mining efficiency relative to the full framework.
• Model-in-the-loop improves both mining and downstream performance. On CAGUI, TP/SR rise from 85.0/64.2 at warm-up to 86.5/67.3 (Stage 1), 87.6/69.1 (Stage 2), and 88.2/69.9 (Stage 3). JudgeAgent's ability to distinguish positive from negative samples trends upward steadily, and MSR and DQA improve markedly after warm-up — with Stage 3 (recycled intents) giving a significant gain in both MSR and DQA.
• Descriptions and preferences help training. Adding descriptions to action data improves TP/SR from 85.2/66.8 to 88.2/69.9; adding preference data (positive samples from trees, negative samples from other branches) further improves to 88.8/70.2.
• Auto-mined data complements public data. With Qwen2.5-VL-7B on CAGUI: no training data gives 74.2 TP / 55.2 SR; public datasets give 84.9 / 64.4; auto-mined data alone gives 87.1 / 69.5; combined public + auto-mined gives 88.8 / 70.2. Auto-mined data alone yields SR +14.3%, CLICK +15.1%, SCROLL +13.4%, TYPE +26.6%, and STOP +17.4%; the combined setting beats public-only by +3.9% TP and +5.8% SR.
• Much lower construction cost. Using assumptions of $7/hour wage, 0.05 h/image annotation, and 0.0014 h/image inspection, the M²-Miner-Agent dataset costs $466 total and $0.02 per image, versus $0.36 per image for Android Control ($31,662), AMEX ($13,680), GUI-Odyssey ($42,816), and AITZ ($6,476). The paper states the method reduces construction cost by $6,010 relative to the similarly sized AITZ and is approximately 18 times more cost-effective per image than all other datasets.
• Higher data quality. In a sample-and-verify check of 100 randomly selected sequences from each of AC, AITZ, and the authors' dataset, the mined-and-inspected data achieves higher Data Quality Accuracy (DQA) than manually annotated datasets.
• Richer intents. A t-SNE visualization of intent text embeddings (Sentence Transformer embeddings reduced with t-SNE) shows that the intent recycling strategy produces a more diverse set of intents from the same initial intent than existing methods.

Methodology in Plain English

The researchers treat data mining as a search problem. Starting from an initial intent and a starting screen state, a tree is grown where each node is a screen (screenshot, a description of the action that led there, an aggregated value estimate, a visit count, and a status of success/failure/intermediate). A search cycles through selection, expansion, simulation, and backpropagation until it finds a trajectory that satisfies the intent.

Three agents make this practical:

• InferAgent proposes K candidate actions for the current screen, using multiple distinct MLLMs, and includes previously generated actions in the prompt to avoid repeats. This raises the hit rate of actions that actually serve the intent.
• OrchestraAgent merges equivalent actions (e.g., taps on the same button at different coordinates) and ranks the rest by how likely they are to achieve the intent, using a multi-choice questioning approach that yields a sorted action queue after K−1 queries. Children are expanded in that order and given decreasing initial UCT values.
• JudgeAgent replaces rollout-based simulation: it judges task completion status directly, assigning a reward of 1 for success and 0 for failure, and for intermediate nodes outputting "valid" or "invalid" whose logits are converted by a softmax into a bounded reward in [0, 1]. Node Q-values are updated with a running weighted average based on visit count.

After mining, the intent recycling step revisits the tree. Every root-to-node path is a candidate trajectory; a dedicated MLLM-based filter scores quality, an MLLM generates a matching new intent for each passing trajectory, and the JudgeAgent verifies that the last node's status is success. Valid pairs are kept without re-mining — for instance, a mis-tap on "book a ride" inside a map app can yield an extra ride-booking trajectory.

To improve mining success rates, the agents are trained progressively in a loop: a warm-up using public datasets, then three stages (basic intents generated from popular apps' home screens with conditional rewriting of time/location; complex intents created by adding conditions and combining functions; recycled intents from the trees), retraining the InferAgent and JudgeAgent with all data mined up to each stage. The whole system runs on a custom six-layer infrastructure (data, engine, algorithm, agent, execution, environment) that unifies execution, mining, and training end to end.

Implementation: Qwen2.5-VL-7B is the base for both InferAgent and JudgeAgent; the OrchestraAgent, the intent recycling filter, and the MLLM generating recycled intents use Qwen2.5-VL-72B. Downstream agents M²-Miner-3B and M²-Miner-7B are trained from Qwen2.5-VL-3B and Qwen2.5-VL-7B. Evaluation uses MSR (mining success ratio), DQA (data quality accuracy), TP (action type prediction accuracy), and SR (step success rate), on AC (Low and High), AITZ, GUI-Odyssey, and CAGUI.

Why This Matters

The paper reframes GUI agent data collection from a labor-bound manual process into an automated, tree-search-driven one, showing that machine-mined data can rival or exceed manually annotated data while being far cheaper. This directly targets the bottleneck the paper identifies: existing GUI agents suffer from limited long-sequence reasoning, poor exception handling, and weak generalization, all traced back to insufficient high-quality GUI data. Its distinctive combination is the coupling of MCTS with cooperative LLM agents and the recognition that failed or off-path branches inside the search tree still contain valuable trajectories.

Real-world applications:

• Automated operation of mobile apps — booking hotels, rides, or flights, and interacting with services that require multi-step form entry.
• Rapidly onboarding new or niche mobile applications where no annotated data exists, using the low per-image mining cost the paper reports.
• Building preference and description-rich training data for reinforcement-learning-style agent training (the paper reports that its data is suitable for RL algorithms, marked "Su.RL: ✓", unlike the compared datasets).
• Assistive automation for users with limited dexterity or vision, who would benefit from an agent that can complete multi-step mobile tasks from a natural-language request.

Industry relevance: The authors are affiliated with Ant Group and Zhejiang University, and the framework is presented as a cost-efficient data-production pipeline (estimated $0.02 per image versus $0.36 for the compared datasets), which is directly relevant to organizations that need to scale GUI agent training across many apps.

Future Directions

• The paper states the work "will be released to facilitate the community research," so broader validation of the framework on additional mobile environments and languages is a natural next step.
• Extending the framework beyond mobile GUIs: the authors note that prior mining work (AgentQ) was limited to parsable web environments, and desktop environments are mentioned as a target of GUI agents generally, but M²-Miner is demonstrated only for mobile.
• Understanding how far the model-in-the-loop loop can be pushed — whether more stages beyond basic, complex, and recycled intents keep improving MSR, DQA, and downstream SR, and how errors compound as intents grow more complex.
• Improving the weakest action categories: the training-data ablation on CAGUI shows SCROLL at 26.6% accuracy and Press at 0.0% across all four data settings, and STOP accuracy actually drops from 78.5% (auto-mined only) to 72.8% (public + auto-mined), indicating open problems in action coverage and balance.
• Reducing reliance on stronger models: the OrchestraAgent, recycling filter, and intent generator use Qwen2.5-VL-72B while the InferAgent and JudgeAgent are Qwen2.5-VL-7B — whether the 72B components can be replaced without losing mining quality is not reported.

Target Audience

Researchers and engineers working on GUI agents, multimodal LLM agents, and agent training-data pipelines; practitioners in industry who need to produce large-scale interaction data for mobile automation without manual annotation; and readers interested in applying MCTS to real-world MLLM-driven search problems. Readers should be comfortable with MCTS terminology, Q-value updates, and agent evaluation metrics such as TP and SR.

Authors’ abstract

Graphical User Interface (GUI) agent is pivotal to advancing intelligent human-computer interaction paradigms. Constructing powerful GUI agents necessitates the large-scale annotation of high-quality user-behavior trajectory data (i.e., intent-trajectory pairs) for training. However, manual annotation methods and current GUI agent data mining approaches typically face three critical challenges: high construction cost, poor data quality, and low data richness. To address these issues, we propose M$^2$-Miner, the first low-cost and automated mobile GUI agent data-mining framework based on Monte Carlo Tree Search (MCTS). For better data mining efficiency and quality, we present a collaborative multi-agent framework, comprising InferAgent, OrchestraAgent, and JudgeAgent for guidance, acceleration, and evaluation. To further enhance the efficiency of mining and enrich intent diversity, we design an intent recycling strategy to extract extra valuable interaction trajectories. Additionally, a progressive model-in-the-loop training strategy is introduced to improve the success rate of data mining. Extensive experiments have demonstrated that the GUI agent fine-tuned using our mined data achieves state-of-the-art performance on several commonly used mobile GUI benchmarks. Our work will be released to facilitate the community research.