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

Faithful Mobile GUI Agents with Guided Advantage Estimator

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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 Faithful-Agent, a two-stage (faithfulness-oriented SFT then GRPO with a guided advantage estimator, GuAE, plus a thought-action consistency reward) training recipe that makes mobile GUI agents abstain under evidence perturbations.

Problem

VLM-based GUI agents often act unfaithfully, relying on memorized shortcuts instead of screen evidence and instructions, and internal reasoning-action inconsistency; standard GRPO suffers advantage collapse under sparse, near-binary GUI rewards.

Approach

Combine SFT that teaches abstention (Back/Home/terminate) under occlusion and instruction-UI mismatch with GRPO modified by GuAE: anchor points {0,1} added to rollout group statistics keep advantages non-zero in constant groups, and a variance-adaptive exponent tempers the advantage scale; plus a rule-based thought-action consistency reward.

Key results

• Faithful-Agent (Stage II) raises Trap SR from 13.88% (Qwen3-VL-8B-Instruct) to 80.21%.
• Stage II achieves 71.67% overall SR versus 41.08% for the Qwen3-VL-8B-Instruct base.
• GRPO+GuAE improves Trap SR by +1.92 over Base GRPO under the same reward (80.21 vs 78.29).

Surprising findings

• Variance-adaptive tempering alone hurts Trap performance (Trap SR drops by 5.62 vs Base GRPO), and only works in combination with the anchor.

Limitations

• Trap data consists exclusively of system-button actions, limiting the diversity of abstention scenarios.
• Trap data relies on automatic annotation with human filtering of UI perturbations, and roughly one third of candidates are discarded.
• Evaluation is on step-wise, offline test splits from two datasets; end-to-end task success is not reported.

Experimental setup

Datasets: AndroidControl AITZ

Benchmarks: AndroidControl AITZ

Metrics: type accuracy grounding accuracy step success rate

Baselines: GUI-Owl-7B UI-TARS-1.5-7B OS-Atlas-Pro-7B Qwen2.5-VL-7B-Instruct Qwen3-VL-8B-Instruct

Builds on: GRPO DAPO GSPO REINFORCE++

Closest papers by meaning

• SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• GUI-CEval: A Hierarchical and Comprehensive Chinese Benchmark for Mobile GUI Agents (CVPR 2026 · similarity 0.88)
• VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics (ICML 2026 · similarity 0.88)
• MMBench-GUI: A Unified Hierarchical Evaluation Framework for Multi-Platform GUI Agents (CVPR 2026 · similarity 0.88)
• UI-Ins: Enhancing GUI Grounding with Multi-Perspective Instruction as Reasoning (ICLR 2026 · similarity 0.87)
• ShowUI: One Vision-Language-Action Model for GUI Visual Agent (CVPR 2025 · similarity 0.87)