Skip to content

M²-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining

Conference: ICLR 2026 arXiv: 2602.05429 Code: Coming soon Area: LLM Agent Keywords: GUI Agent, MCTS, Data Mining, Multi-Agent Collaboration, Mobile Interaction

TL;DR

This paper proposes M²-Miner, the first MCTS-based automatic data mining framework for mobile GUI agents. Through the collaboration of three agents—InferAgent, OrchestraAgent, and JudgeAgent—it achieves a 64× improvement in mining efficiency, enriches intent diversity via an intent recycling strategy, and trains a GUI agent that achieves state-of-the-art performance on multiple benchmarks.

Background & Motivation

Background: GUI agents automate software interactions by interpreting user intents and executing action sequences on graphical interfaces, representing a prominent research direction in both academia and industry. High-quality intent–trajectory training data is the core dependency of current GUI agents.

Limitations of Prior Work: - High cost: Manual annotation (e.g., AITW, AndroidControl) requires hours per sample, with costs as high as $0.36/image. - Low quality: Both manually annotated and automatically mined data frequently contain redundant steps, ambiguous intent descriptions, and biased action paths. - Low diversity: Existing datasets adopt an intent-to-flat-trajectory structure, recording only a single successful path per intent, resulting in monotonic intent types.

Key Challenge: Manual annotation is quality-controllable but not scalable, while existing automated mining methods (e.g., AgentQ based on vanilla MCTS, OS-Genesis based on rule-driven exploration) suffer from low efficiency, are limited to web environments, or produce only single trajectories.

Goal: How to automatically mine high-quality, high-diversity mobile GUI interaction trajectory data at low cost?

Key Insight: The paper introduces MCTS into mobile GUI data mining, but vanilla MCTS suffers from extremely low efficiency due to random expansion. The authors observe that: (a) the expansion phase requires intelligent guidance rather than random exploration; (b) the simulation phase can replace rollouts with process rewards; (c) non-primary paths in the search tree contain additional valuable intent–trajectory pairs.

Core Idea: MCTS + three-agent collaboration (guided expansion + accelerated ranking + process evaluation) + intent recycling = efficient, high-quality, and high-diversity GUI data mining.

Method

Overall Architecture

M²-Miner uses MCTS as its backbone. Given an initial intent \(I_0\) and a starting GUI state \(s_0\), it outputs an intent–trajectory tree \(\mathcal{T}=(\mathcal{V},\mathcal{A},\mathcal{P},\mathcal{I})\) containing valid interaction trajectories \(\tau=(s_0,a_0,s_1,\ldots)\). Each node in the tree stores a screenshot, action description, Q-value, visit count, and task completion status. The four MCTS phases (selection → expansion → simulation → backpropagation) iterate until a trajectory matching the intent is mined.

Key Designs

  1. InferAgent (Expansion Phase — Action Generation):

    • Function: Generates \(K\) candidate actions for the selected node.
    • Mechanism: Reasons about the most likely correct action based on the current GUI screenshot and target intent. Multiple different MLLMs are used for generation to ensure diversity of the action space; previously generated actions are included in the prompt to avoid duplicates.
    • Design Motivation: Replaces the random expansion of vanilla MCTS, substantially increasing the hit rate of correct actions.

  2. OrchestraAgent (Expansion Phase — Action Ranking and Deduplication):

    • Function: Merges semantically equivalent actions (e.g., clicks on the same button at different coordinates) and ranks them by their likelihood of achieving the target intent.
    • Mechanism: An MLLM selects the most promising action at each iteration via a multi-choice question format; \(K-1\) queries yield a ranked queue. Ranked actions are assigned decreasing initial UCT values in order.
    • Design Motivation: Avoids redundant expansion and ensures that the search prioritizes the most promising branches.

  3. JudgeAgent (Simulation Phase — Process Reward Estimation):

    • Function: Analyzes the GUI screenshot of newly expanded nodes, assesses task completion status, and computes rewards.
    • Mechanism: Terminal nodes (success/failure) receive rewards of 1/0. For intermediate nodes, the MLLM head outputs logits for "valid"/"invalid", normalized via softmax into a \([0,1]\) probability as the reward: \(r_{\text{intermediate}} = \frac{\exp(logits_{\text{valid}})}{\exp(logits_{\text{valid}}) + \exp(logits_{\text{invalid}})}\)
    • Design Motivation: Replaces the full rollout required by vanilla MCTS to evaluate rewards. Node Q-values are updated as: \(Q_i = \frac{Q_{i-1} \times N_{i-1} + R_i}{N_{i-1}+1}\), greatly reducing computational overhead in the simulation phase.

  4. Intent Recycling Strategy:

    • Function: Extracts additional intent–trajectory pairs from non-primary paths in a completed mining tree.
    • Mechanism: All root-to-node paths in the tree are enumerated and evaluated by an intent recycling filter (implemented via an MLLM); paths that pass are assigned new intents generated by an MLLM, which are then verified for consistency with the trajectory by JudgeAgent.
    • Design Motivation: Evolves the structure from "one tree, one intent" to "one tree, multiple intents," yielding more diverse data without re-mining. For example, when mining the intent "query a route," an accidental tap on a "ride-hailing" button may produce a valid ride-hailing trajectory.

  5. Progressive Model-in-the-Loop Training Strategy:

    • Function: Iteratively improves the capabilities of InferAgent and JudgeAgent.
    • Mechanism: Three progressive training stages — Stage 1: basic intents (common services + conditional rewriting) → Stage 2: complex intents (function combinations + retry on failure) → Stage 3: recycled intents (applying intent recycling to historical trees). Data mined at each stage is used for continual training of both agents.
    • Design Motivation: Forms a positive feedback loop in which agent capability and data complexity grow in tandem.

Loss & Training

  • InferAgent and JudgeAgent are fine-tuned based on Qwen2.5-VL-7B.
  • OrchestraAgent and the intent recycling filter use Qwen2.5-VL-72B.
  • Training data also includes description information and preference data (constructed from positive/negative paths).

Key Experimental Results

Main Results

Model AC-Low TP/SR AC-High TP/SR AITZ TP/SR GUI-Odyssey TP/SR CAGUI TP/SR
GPT-4o 74.3/19.4 66.3/20.8 70.0/35.3 - 3.67/3.67
UI-TARS-7B* 98.0/90.8 83.7/72.5 80.4/65.8 90.1/87.0 88.6/70.0
OS-Genesis-7B 90.7/74.2 66.2/44.5 20.0/8.5 11.7/3.6 38.1/14.5
GUI-Owl-7B 93.8/90.0 81.5/72.8 78.9/65.1 83.4/60.7 80.0/59.2
M²-Miner-7B 97.5/93.5 81.8/72.9 81.3/69.4 90.5/79.3 88.8/70.2

*UI-TARS-7B uses large-scale private manually annotated data.

Ablation Study

Configuration TP SR Note
Warm-up 85.0 64.2 Pre-trained on public data only
+ Stage 1 (Basic Intents) 86.5 67.3 +3.1% SR
+ Stage 2 (Complex Intents) 87.6 69.1 +1.8% SR
+ Stage 3 (Recycled Intents) 88.2 69.9 +0.8% SR, cumulative +5.7%
Act only 85.2 66.8 Action labels only
Act + Description 88.2 69.9 +3.1% SR
Act + Des + Preference 88.8 70.2 +3.4% SR

Key Findings

  • Exponential efficiency gains: Compared to vanilla MCTS, M²-Miner achieves a 64× efficiency improvement at a task length of 9. OrchestraAgent reduces redundant nodes in the expansion phase; JudgeAgent eliminates rollouts in the simulation phase.
  • 18× cost reduction: The M²-Miner-Agent dataset costs only $0.02 per image, versus $0.36 for manually annotated datasets.
  • Higher data quality: Manual inspection of 100 randomly sampled records shows that M²-Miner's data quality accuracy (DQA) surpasses that of manually annotated datasets (AC and AITZ).
  • Description and preference data are beneficial: Compared to action labels alone, adding description information improves SR by 3.1%; adding preference data yields a further 0.3% gain.
  • Strong generalization to unseen scenarios: On CAGUI, where no training data is available, Qwen2.5-VL-7B trained on M²-Miner data improves SR from 55.2% to 70.2%.

Highlights & Insights

  • The MCTS + multi-agent paradigm is elegantly designed: The three agents each fulfill a distinct role—generation, ranking, and evaluation—precisely addressing the three bottlenecks of vanilla MCTS in the GUI domain (random expansion, redundant nodes, and expensive rollouts).
  • Intent Recycling is a highly creative design: It repurposes "failed" exploratory paths as an additional data source, simultaneously addressing intent diversity and mining efficiency. This idea is transferable to other agent data collection settings.
  • Process rewards as rollout substitutes: Using MLLM logit probabilities as intermediate rewards preserves the theoretical advantages of MCTS while avoiding the high cost of full simulation. This design has reference value for all MCTS+LLM systems.

Limitations & Future Work

  • The framework is validated only on mobile platforms; applicability to desktop and web environments remains unexplored.
  • OrchestraAgent relies on a 72B model (Qwen2.5-VL-72B), incurring high deployment costs.
  • The quality of intent recycling filtering and intent generation depends on MLLM capability and may degrade for complex applications.
  • The dataset scale (20k images, 2,565 trajectories) still lags behind the private data used by UI-TARS.
  • The three-stage model-in-the-loop training requires multiple mining → training cycles; the practical engineering cost is not reported in detail.
  • vs. AgentQ: AgentQ pioneered MCTS-based data mining for web environments but is limited to parseable HTML environments and suffers from low efficiency. M²-Miner extends the approach to vision-driven mobile platforms, improving efficiency by orders of magnitude through multi-agent collaboration.
  • vs. OS-Genesis: OS-Genesis employs unsupervised rule-based interaction and reverse task synthesis, requiring no predefined tasks but yielding lower data quality (SR of only 8.5% on AITZ). M²-Miner's structured MCTS search produces higher-quality trajectories.
  • vs. UI-TARS: UI-TARS relies on large-scale private manually annotated data, yet M²-Miner surpasses it on nearly all SR metrics, demonstrating that the potential of automated mining frameworks has exceeded that of manual annotation.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ The combination of MCTS, multi-agent collaboration, and intent recycling is pioneering in GUI data mining, with clear technical contributions from each component.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Five benchmarks, multiple ablation groups (multi-agent design, training strategy, data structure, training data), efficiency analysis, and cost comparison — highly comprehensive.
  • Writing Quality: ⭐⭐⭐⭐ Well-structured with rich illustrations, though some content is repetitive and the paper is lengthy.
  • Value: ⭐⭐⭐⭐⭐ Provides a highly valuable data production paradigm for the GUI agent community, with empirical evidence that automated mining can surpass manual annotation.