Research map / Multi-agent systems and games / LLM agents

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

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes M²-Miner, an MCTS-based automated framework for mining mobile GUI intent-trajectory data using three collaborative agents, intent recycling, and progressive model-in-the-loop training, yielding a 2,565-trajectory dataset and Qwen2.5-VL agents that report SOTA on several mobile GUI benchmarks.

Problem

Manual annotation of GUI intent-trajectory data is costly, and existing automated mining methods yield low-quality, flat, and low-diversity data, limiting GUI agent training.

Approach

Applies MCTS to mobile GUI data mining with a multi-agent setup (InferAgent proposes actions, OrchestraAgent merges and ranks them, JudgeAgent gives process rewards instead of rollouts), recycles intents from non-primary tree paths, and iteratively retrains the agents in the loop.

Key results

• M²-Miner-7B reaches 93.5% SR on AC-Low, 69.4% SR on AITZ (TP 81.3%), and 70.2% SR on CAGUI, up from 55.2% for the Qwen2.5-VL-7B base.
• The multi-agent framework gives a 64x mining efficiency speedup over vanilla MCTS at task length 9.
• Stage-wise model-in-the-loop training raises CAGUI SR from 64.2 (warm-up) to 69.9 (stage 3).

Surprising findings

• Using only auto-mined data on CAGUI gives a larger SR gain (+14.3%) than public datasets (+9.2%).

Limitations

• Relies on large MLLMs (Qwen2.5-VL-72B) for orchestration, filtering and intent generation, so mining cost depends on them.
• Dataset quality check uses only minimal manual inspection; quality estimates rest on 100-sample checks per dataset.

Experimental setup

Datasets: AndroidControl AITZ GUI Odyssey AMEX CAGUI

Benchmarks: AndroidControl-Low AndroidControl-High AITZ GUI-Odyssey CAGUI

Metrics: TP (action type accuracy) SR (step success rate) MSR (mining success ratio) DQA (data quality accuracy)

Baselines: GPT-4o Gemini 2.0 Claude UI-TARS-7B OdysseyAgent SpiritSight-8B GUI-R1-7B OS-Atlas-7B Aguvis-7B InfiGUI-R1 Qwen2.5-VL-3B Qwen2.5-VL-7B

Builds on: Monte Carlo Tree Search UCT AgentQ OS-Genesis Mobile-Agent-v3

Closest papers by meaning

• TongUI: Internet-Scale Trajectories from Multimodal Web Tutorials for Generalized GUI Agents (AAAI 2026 · similarity 0.88)
• UI-Genie: A Self-Improving Approach for Iteratively Boosting MLLM-based Mobile GUI Agents (NeurIPS 2025 · similarity 0.87)
• VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics (ICML 2026 · similarity 0.87)
• Video2GUI: Synthesizing Large-Scale Interaction Trajectories for Generalized GUI Agent Pretraining (ICML 2026 · similarity 0.86)
• MMBench-GUI: A Unified Hierarchical Evaluation Framework for Multi-Platform GUI Agents (CVPR 2026 · similarity 0.86)
• GUI-Xplore: Empowering Generalizable GUI Agents with One Exploration (CVPR 2025 · similarity 0.86)