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

K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration

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 K²-Agent, a hierarchical mobile-control agent that pairs a training-free VLM planner evolving declarative task knowledge via an SRLR loop with a Qwen2.5-VL-7B executor trained by curriculum-guided GRPO (C-GRPO), reaching 76.1% on AndroidWorld from screenshots only.

Problem

Mobile device control agents fail on long-horizon, precise multi-step tasks; existing systems are either training-free (capped by closed foundation models) or learning-based monolithic policies that conflate planning and execution, requiring large data and compute.

Approach

Separates declarative ("know-what") knowledge, evolved non-parametrically by a Summarize-Reflect-Locate-Revise loop from one demonstration, from procedural ("know-how") skills, learned by a 7B executor with C-GRPO that balances replay pools by decoupled type/parameter errors and injects variable-length expert action prefixes scheduled by difficulty and training stage.

Key results

• K²-Agent reaches 76.1 ± 1.0% success on AndroidWorld from screenshots only, above UI-TARS-2 and Mobile-Agent-v3 (73.3%).
• Zero-shot executor achieves 91.3% overall on ScreenSpot-v2, below UI-Venus-Ground-72B (95.3%) but above Claude 3.7 Sonnet (87.6%).
• Ablation: no hierarchy 35.3%, no hierarchy + SRLR 58.6%, SRLR + SFT-low 62.0%, SRLR + GRPO-low 68.9%, full 76.1%.

Surprising findings

• Full K²-Agent is stated to beat several closed-source screenshot-only agents while the executor is only 7B; the ScreenSpot-v2 executor does not beat the best specialized grounding model.

Limitations

• Results on AndroidWorld averaged over only 3 runs with ±1.0 std; comparisons mix screenshot-only and A11y-based baselines.
• Executor underperforms the best specialized grounding model on ScreenSpot-v2 mobile icon accuracy (80.6 vs 93.8 for UI-Venus).

Experimental setup

Datasets: AndroidWorld demonstrations (self-collected) AndroidWorld AndroidInTheWild ScreenSpot-v2

Benchmarks: AndroidWorld ScreenSpot-v2 AndroidInTheWild (General, WebShopping)

Metrics: success rate accuracy

Baselines: M3A AndroidGen Agent S2 MobileUse DroidRun FinalRun InfiGUI Agent GUI-critic-R1 UI-TARS UI-Venus Seed1.5-VL Mobile-Agent-v3

Builds on: GRPO DPO Qwen2.5-VL

Closest papers by meaning

• Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation (AAAI 2026 · similarity 0.88)
• Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration (NeurIPS 2024 · similarity 0.86)
• ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.86)
• MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models (NeurIPS 2026 · similarity 0.86)
• DistRL: An Asynchronous Distributed Reinforcement Learning Framework for On-Device Control Agent (ICLR 2025 · similarity 0.86)
• Agent S: An Open Agentic Framework that Uses Computers Like a Human (ICLR 2025 · similarity 0.86)