UI-R1: Enhancing Efficient Action Prediction of GUI Agents by Reinforcement Learning

DeepSeek‑R1 proved that reinforcement learning can coax reasoning out of large language models, but it never left the text world. The authors therefore set out to test whether the same rule‑based RL tricks work for multimodal models that must click, swipe, and type on a screen. Their target is to raise the success rate of GUI‑action prediction without inflating model size.

UI‑R1 adds a rule‑based action reward that scores each predicted GUI operation against a handcrafted correctness table. The reward feeds a policy optimizer, Group Relative Policy Optimization, so the model learns to prefer actions that satisfy the rules. This mirrors the language‑task setup but swaps token‑level feedback for GUI‑specific correctness signals.

To keep inference fast, UI‑R1‑Efficient splits training into two stages. First, the model learns the full reasoning chain; second, a distilled version learns to truncate that chain while preserving the high‑reward decisions. The authors also built a 2 000‑example dataset covering five common mobile actions, ensuring the RL signal sees diverse, challenging scenarios.

Experiments report that UI‑R1‑3B lifts average accuracy by 18.3 % on ScreenSpot, 6.0 % on ScreenSpot‑Pro, and 10.9 % on ANDROIDCONTROL compared with the base Qwen2.5‑VL‑3B. The efficient variant matches the performance of much larger state‑of‑the‑art models while using fewer parameters, confirming that rule‑based RL can close the gap between size and skill.

If the reward design scales to richer interaction vocabularies, developers could train lightweight agents that still reason about complex UI flows. The abstract leaves open how the approach handles dynamic content or real‑time latency, questions that future work will need to address.

TakeawayRule‑based reinforcement learning can substantially improve GUI action prediction without growing model size.