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

Scaling Synthetic Task Generation for Agents via Exploration

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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 AutoPlay, an MLLM explorer-plus-task-generator pipeline that synthesizes diverse, feasible, verifiable UI agent tasks (20k Android, 10k Ubuntu) for SFT and GRPO-based RL of Qwen2.5-VL agents.

Problem

Post-training MLLM UI agents requires large sets of diverse, feasible, and verifiable downstream tasks, which are costly to get via human annotation and poorly covered by prompting MLLMs without environment grounding.

Approach

An MLLM explorer with episodic memory (summarized prior trajectories) first explores the app environment goal-agnostically; a task generator then conditions on these exploration trajectories plus task guideline prompts to propose environment-grounded tasks, which are executed by a GPT-4o-based executor and filtered by an MLLM verifier.

Key results

• AutoPlay-7B reaches 40.1% Pass@1 on AndroidWorld, +20.6 over Qwen2.5-VL-7B (19.5%).
• AutoPlay-72B reaches 14.5% Pass@1 on OSWorld, +10.1 over Qwen2.5-VL-72B (4.4%).
• RL with the verifier lifts AutoPlay-3B from 34.2% to 39.9% Pass@1 on AndroidWorld (+5.7).

Surprising findings

• Qwen2.5-VL-72B as verifier yields only a 1.5-point drop vs GPT-4o (38.6 vs 40.1), suggesting open models can serve as verifiers.
• Iterative Exploration has a higher executor success rate (56.4%) than AutoPlay (46.0%) yet trains worse agents.

Limitations

• Computer-use task guidelines lack diversity, so cross-app, bash, and web search tasks are under-covered on OSWorld.
• Executor relies on GPT-4o and grounding models, and RL is restricted to tasks the executor solved at least once.
• Improvements on OSWorld remain well below UI-TARS, which uses human-labeled GUI data.

Experimental setup

Datasets: AutoPlay-generated Android tasks AutoPlay-generated Ubuntu tasks

Benchmarks: AndroidWorld OSWorld

Metrics: Pass@1 Pass@5 task execution success rate

Baselines: PAE No Exploration Iterative Exploration UI-TARS-1.5 7B UI-TARS-1.5 72B UI-TARS 2 Seed-VL 1.5 AguVis 72B GPT-4o + UI-TARS o3 + GTA1 Qwen2.5-VL 7B Qwen2.5-VL 72B

Builds on: GRPO PAE Qwen2.5-VL AndroidWorld OSWorld

Closest papers by meaning

• ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training (ICML 2026 · similarity 0.88)
• AgentSynth: Scalable Task Generation for Generalist Computer-Use Agents (ICLR 2026 · similarity 0.88)
• OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (NeurIPS 2024 · similarity 0.88)
• Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.87)
• Exploration-Driven Generative Interactive Environments (CVPR 2025 · similarity 0.87)
• Scaling Autonomous Agents via Automatic Reward Modeling And Planning (ICLR 2025 · similarity 0.87)