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

CoME: Empowering Channel-of-Mobile-Experts with Informative Hybrid-Capabilities Reasoning

ICML 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 CoME, a mobile-agent MoE-like architecture with four stage-aligned experts selected by output-oriented activation, trained via Expert-FT, Router-FT, CoT-FT and an information-gain-driven DPO (Info-DPO), reporting state-of-the-art action accuracy on AITZ and AMEX.

Problem

Mobile agents need hybrid-capabilities reasoning (screen summary, subtask planning, action decision, action function), but existing agents either enhance single capabilities without integration or pretrain broadly with unbalanced performance, and MoE's input-oriented routing cannot align experts wit

Approach

Four FFN experts, one per reasoning stage, share self-attention; a learned channel router fuses expert channel states per output token (output-oriented activation), trained in three stages, plus Info-DPO that scores each reasoning stage by the information gain it gives the final action and uses it to build DPO pairs.

Key results

• CoME (5B) reaches 66.98% action match on AITZ, versus 65.41% for UI-TARS-7B and 63.56% for AriaUI, the best overall in the reported table.
• On AMEX, CoME achieves 72.61% overall, +1.90 over the dense Qwen2-VL-7B-based result and +8.05 over the sparse MoE baseline as stated in the paper.
• Ablation: removing Info-DPO drops AITZ match from 66.98 to 62.93, and removing Router-FT drops it to 60.05.

Surprising findings

• A 7B reward model for InfoGain performs comparably to a 72B reward model, with the 72B model yielding smaller InfoGain differences and weaker distinguishing ability.

Limitations

• MoE-style comparison baselines were fine-tuned by the authors on mobile data, and the efficiency table reports only one training and inference configuration.
• Gains over strong 7B baselines on AITZ are modest (about +1.6 points over UI-TARS-7B), and some action categories (e.g., PRESS, TYPE on AMEX) underperform baselines.

Experimental setup

Datasets: AITZ AMEX

Benchmarks: AITZ AMEX

Metrics: action type accuracy action match accuracy

Baselines: Auto-GUI SeeClick ShowUI UI-TARS UGround OS-Atlas AriaUI GUI-R1 UI-S1 MolmoE DeepSeek-VL2 SphAgent

Builds on: Mixture-of-Experts DPO Qwen2-VL chain-of-action-thought

Closest papers by meaning

• Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration (NeurIPS 2024 · similarity 0.84)
• ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.84)
• K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control (ICLR 2026 · similarity 0.84)
• Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation (AAAI 2026 · similarity 0.83)
• MobileIPL: Enhancing Mobile Agents Thinking Process via Iterative Preference Learning (ICLR 2026 · similarity 0.83)
• InfoPO: Information-Driven Policy Optimization for User-Centric Agents (ICML 2026 · similarity 0.83)