Qwen-Planner-Agent tops MobilePA-Bench via closed-loop framework

"Qwen-Planner-Agent achieves the best overall performance among all evaluated models and systems on MobilePA-Bench," a new arXiv preprint submitted 24 September 2026 reports. The abstract publishes no absolute scores or per-competitor numbers to size that claim.

The 26-author team, led by Tingyu Qu, frames the paper around whether AI can be "both the object of development and an active participant in building next-generation AI systems." Their answer is a closed-loop pipeline with three stages. "AI for Data" runs a "human-gated agentic data flywheel" where specialized agents build tasks, collect interaction trajectories, curate training data and use training feedback to guide the next round of data generation. "AI for Training" combines a supervised planning cold start with online agentic reinforcement learning under a method the authors call Competence-Aware Reward-and-Advantage Engineering (CARE), meant to "reduce reasoning and tool-use costs while preserving task performance." A third stage drives what the paper describes as model-harness co-evolution, an execution-evidence-driven loop that feeds structured action feedback and preserved failure traces back into both the model and the runtime harness.

The authors call mobile planning a demanding test because "complex, long-horizon tasks challenge agent reliability, while costly real-device interaction limits development scalability." They also report "improvements across non-mobile agentic benchmarks while largely preserving general capabilities," without publishing figures for those transfer claims in the abstract.