Skip to content

M\(^2\)-Miner: Multi-Agent Enhanced MCTS for Mobile GUI Agent Data Mining

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

MCTS-guided multi-agent framework for scalable high-quality GUI trajectory data mining

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

M²-Miner is an automated mobile GUI agent data-mining framework that replaces expensive manual annotation by applying Monte Carlo Tree Search (MCTS) augmented with a three-agent collaboration (InferAgent, OrchestraAgent, JudgeAgent). An intent recycling strategy extracts additional intent-trajectory pairs from non-primary tree paths, and a progressive model-in-the-loop training scheme iteratively improves agent capabilities across three curriculum stages. GUI agents fine-tuned on the resulting dataset achieve state-of-the-art performance on mobile benchmarks.

Problem

Training GUI agents requires large-scale, high-quality intent-to-trajectory data, but existing construction methods suffer from three compounding failures: (1) manual annotation costs several hours per entry; (2) automated and human datasets contain redundant steps, ambiguous intents, and biased paths; and (3) datasets record only a single successful path per intent (flat structure), limiting richness and diversity. Vanilla MCTS applied naïvely to this domain compounds the problem with random expansion and slow rollout-based reward computation.

Method

M²-Miner formalizes mobile GUI exploration as intent-trajectory trees \(\mathcal{T} = (V, A, P, I)\), where each node \(v = (\text{img}_v, \text{meta}_v, Q_v, N_v, \text{stat}_v)\) stores a screenshot, the action that reached it, a Q-value, a visit count, and a terminal status.

Collaborative multi-agent MCTS: - InferAgent (expansion): generates \(K\) candidate actions from the current screenshot using multiple MLLMs; prevents re-generation of prior actions. - OrchestraAgent (expansion): deduplicates equivalent actions and ranks survivors via \(K-1\) multi-choice queries, assigning decreasing initial UCT scores to child nodes. - JudgeAgent (simulation): replaces rollout-based rewards with process-based rewards. For terminal nodes reward \(\in \{0,1\}\); for intermediate nodes it uses MLLM logits:

\[r_{\text{intermediate}} = \frac{\exp(\text{logits}_{\text{valid}})}{\exp(\text{logits}_{\text{valid}}) + \exp(\text{logits}_{\text{invalid}})}\]

Q-values update online:

\[Q_i = \frac{Q_{i-1} \times N_{i-1} + R_i}{N_{i-1} + 1}, \quad N_i = N_{i-1} + 1\]

Intent recycling: after each tree is finished, every root-to-node path is scored by a recycling filter; passing paths get new intents generated by an MLLM and verified by JudgeAgent, yielding bonus intent-trajectory pairs without re-mining.

Progressive model-in-the-loop training: three curriculum stages — basic intents (single-app tasks with conditional rewrites), complex intents (multi-step combinations and previously failed intents), and recycled intents — each followed by retraining InferAgent and JudgeAgent on all accumulated data.

Key Contributions

  • First MCTS-based automated mobile GUI agent data-mining framework (M²-Miner), replacing flat trajectory collection with multi-intent tree structures.
  • Collaborative three-agent architecture (InferAgent / OrchestraAgent / JudgeAgent) that replaces random expansion and rollout simulation, achieving exponential efficiency gains.
  • Intent recycling strategy that extracts additional intent-trajectory pairs from non-primary tree paths, improving both efficiency and intent diversity.
  • Progressive model-in-the-loop training curriculum (basic → complex → recycled) that bootstraps mining capability from public data and continuously self-improves.
  • M²-Miner-Agent dataset: 2,565 trajectories (avg. length 7.8) over 20k images, with richer descriptive metadata than flat alternatives.

Results

  • Mining efficiency: \(64\times\) improvement over vanilla MCTS at task length 9.
  • Downstream benchmark: GUI agents fine-tuned on M²-Miner data achieve SOTA on several commonly used mobile GUI benchmarks (specific metric names/numbers truncated from the provided text).
  • Intent diversity: t-SNE visualization of Sentence Transformer embeddings shows broader coverage than prior methods under the same initial intents.
  • Ablation validates contribution of each component: multi-agent framework, intent recycling, and progressive training all independently improve success rate and data quality.
  • Trajectory data enriched with step descriptions (meta annotations) yields greater training benefit than traditional flat datasets.

Limitations

  • Dataset scale is modest (2,565 trajectories), and the paper acknowledges minimal manual quality inspection is still required.
  • The system depends on MLLM quality for InferAgent and JudgeAgent; weak base models produce many invalid explorations and inaccurate termination judgments, motivating the warm-up stage but not eliminating the dependency.
  • Evaluation is confined to mobile GUI environments; applicability to desktop or web-only scenarios (non-vision) is not demonstrated.
  • Intent recycling filter and new-intent generation rely on MLLM accuracy; misaligned generated intents are discarded but represent wasted compute.
  • The paper text provided is truncated, so precise benchmark numbers and full comparison tables are unavailable for citation here.

Relevance to Agentic AI / LLM Agents

M²-Miner directly addresses the data bottleneck that constrains GUI agents — a canonical instantiation of the broader problem of scalably generating high-quality agentic training data without human labelers. The multi-agent orchestration pattern (specialized roles for inference, action ranking, and reward estimation) is a transferable blueprint for any MCTS-augmented agentic pipeline. The intent recycling strategy is an instance of self-distillation: reusing exploration byproducts as supervision signal, which resonates with recent self-improvement loops (STaR, RLVR). For researchers tracking LLM agents, this work situates squarely at the intersection of automated data synthesis, tool-use agents, and curriculum training — themes central to advancing autonomous agents beyond prompt-only regimes.