{"paper":{"title":"UI-R1: Enhancing Efficient Action Prediction of GUI Agents by Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Rule-based RL on 136 GUI tasks lifts a 3B multimodal model to 22% higher action-prediction accuracy.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Guanjing Xiong, Han Xiao, Hao Wang, Hongsheng Li, Liang Liu, Shuai Ren, Xi Yin, Yaxuan Guo, Yuxiang Chai, Zhengxi Lu","submitted_at":"2025-03-27T15:39:30Z","abstract_excerpt":"The recent DeepSeek-R1 has showcased the emergence of reasoning capabilities in LLMs through reinforcement learning (RL) with rule-based rewards. Despite its success in language models, its application in multi-modal domains, particularly in graphic user interface (GUI) agent tasks, remains under-explored. To address this issue, we propose UI-R1, the first framework to explore how rule-based RL can enhance the reasoning capabilities of multimodal large language models (MLLMs) for GUI action prediction tasks. Specifically, UI-R1 introduces a novel rule-based action reward, enabling model optimi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"UI-R1-3B achieves significant improvements over the base model (Qwen2.5-VL-3B) on both in-domain and out-of-domain tasks, with average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The rule-based action reward provides sufficient and unbiased supervision for policy optimization across diverse GUI tasks without post-hoc adjustments or hidden data selection.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"UI-R1 shows rule-based RL with GRPO on 136 GUI tasks improves a 3B MLLM's action prediction accuracy by 6-22% over its base model and matches larger SFT-trained models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Rule-based RL on 136 GUI tasks lifts a 3B multimodal model to 22% higher action-prediction accuracy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e3caecae84a458b7a808b92251183a3314729b6427e483a02ffd817da673bad4"},"source":{"id":"2503.21620","kind":"arxiv","version":5},"verdict":{"id":"30738ba4-462a-40ec-b816-9585f8013006","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T10:58:22.234334Z","strongest_claim":"UI-R1-3B achieves significant improvements over the base model (Qwen2.5-VL-3B) on both in-domain and out-of-domain tasks, with average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL.","one_line_summary":"UI-R1 shows rule-based RL with GRPO on 136 GUI tasks improves a 3B MLLM's action prediction accuracy by 6-22% over its base model and matches larger SFT-trained models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The rule-based action reward provides sufficient and unbiased supervision for policy optimization across diverse GUI tasks without post-hoc adjustments or hidden data selection.","pith_extraction_headline":"Rule-based RL on 136 GUI tasks lifts a 3B multimodal model to 22% higher action-prediction accuracy."},"references":{"count":18,"sample":[{"doi":"","year":null,"title":"L1: Controlling how long a reasoning model thinks with reinforcement learning","work_id":"ad7236fb-3752-48a9-b782-a384899c45a0","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"arXiv preprint arXiv:2407.17490 , year=","work_id":"07ffe67e-3e5a-406c-84fe-3de84f8dd21d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"VisRL: Intention-driven visual perception via reinforced reasoning","work_id":"5e295e70-d8e8-4932-8908-266167a5615d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents","work_id":"9def1724-6fd2-4d5b-8339-4c1ee76e62f8","ref_index":4,"cited_arxiv_id":"2410.05243","is_internal_anchor":true},{"doi":"","year":null,"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","ref_index":5,"cited_arxiv_id":"2501.12948","is_internal_anchor":true}],"resolved_work":18,"snapshot_sha256":"a649709e89f6b39e4fb656d1006ad44a48abaa31c613a19dacd1763fd7b79282","internal_anchors":9},"formal_canon":{"evidence_count":2,"snapshot_sha256":"55ea065743025a6f1c3bb76e55ec151e339d60888732c45d4e54d91087d161b8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}