pith. machine review for the scientific record. sign in

arxiv: 2503.21620 · v5 · submitted 2025-03-27 · 💻 cs.AI

Recognition: 2 theorem links

· Lean Theorem

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

Zhengxi Lu , Yuxiang Chai , Yaxuan Guo , Xi Yin , Liang Liu , Hao Wang , Han Xiao , Shuai Ren
show 2 more authors
Guanjing Xiong Hongsheng Li
Authors on Pith no claims yet

Pith reviewed 2026-05-16 10:58 UTC · model grok-4.3

classification 💻 cs.AI
keywords GUI agentsreinforcement learningmultimodal modelsaction predictionrule-based rewardsGRPOscreen groundingmobile interfaces
0
0 comments X

The pith

Rule-based RL on 136 GUI tasks lifts a 3B multimodal model to 22% higher action-prediction accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes UI-R1 as the first application of rule-based reinforcement learning to multimodal models for GUI action prediction. It introduces a rule-based action reward that scores outputs for use in GRPO optimization, trained on a small curated set of 136 mobile-device tasks spanning five action types. This produces UI-R1-3B, which records average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL over the Qwen2.5-VL-3B base model, while matching or exceeding larger models trained by supervised fine-tuning on 76K samples. A reader cares because the result indicates that high-quality rule-based signals can replace large labeled datasets for building capable GUI agents.

Core claim

UI-R1 is the first framework to apply rule-based RL to GUI action prediction by defining a rule-based action reward that evaluates predicted actions against ground truth, then using this reward inside Group Relative Policy Optimization to train a 3B multimodal model on only 136 tasks; the resulting UI-R1-3B model delivers the stated accuracy gains on in-domain and out-of-domain benchmarks and remains competitive with 7B-scale models trained via supervised fine-tuning on far larger corpora, while the optimized UI-R1-E-3B variant further raises both accuracy and grounding efficiency.

What carries the argument

The rule-based action reward that supplies scalar feedback on action correctness for GRPO policy updates.

If this is right

  • UI-R1-3B improves both in-domain and out-of-domain GUI grounding and control.
  • The optimized UI-R1-E-3B variant raises grounding efficiency while preserving accuracy gains.
  • Rule-based RL produces competitive GUI agents with far less data than supervised fine-tuning on 76K samples.
  • The same reward-plus-GRPO pattern can be reused for other common mobile action types without additional hand-labeled data.
  • Performance remains stable across the five action categories covered in the 136-task set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same rule-based reward design could be extended to multi-step planning agents that chain several GUI actions.
  • Because only 136 tasks suffice, the method suggests that targeted task selection may matter more than dataset scale for GUI agents.
  • The efficiency gains in the E variant imply that reward shaping can also be used to optimize inference speed in addition to accuracy.
  • The approach may transfer to non-mobile GUI environments if equivalent rule-based correctness checks can be defined.

Load-bearing premise

The rule-based action reward supplies accurate and unbiased supervision for every task type and environment encountered during training.

What would settle it

Retraining the same 3B model with the identical dataset and GRPO but replacing the rule-based reward with a constant or random reward, then observing that the accuracy gains on ScreenSpot and ANDROIDCONTROL disappear or reverse.

read the original abstract

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 optimization via policy-based algorithms such as Group Relative Policy Optimization (GRPO). For efficient training, we curate a small yet high-quality dataset of 136 challenging tasks, encompassing five common action types on mobile devices. Experimental results demonstrate that our proposed UI-R1-3B achieves significant improvements over the base model (i.e. Qwen2.5-VL-3B) on both in-domain (ID) and out-of-domain (OOD) tasks, with average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL. Furthermore, UI-R1-3B delivers competitive performance compared to larger models (e.g., OS-Atlas-7B) trained via supervised fine-tuning (SFT) on 76K samples. We additionally develop an optimized version, UI-R1-E-3B, which significantly improves both grounding efficiency and accuracy. These results underscore the potential of rule-based reinforcement learning to advance GUI understanding and control, paving the way for future research in this domain. Code website: https://github.com/lll6gg/UI-R1.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes UI-R1, the first framework to apply rule-based reinforcement learning with Group Relative Policy Optimization (GRPO) to multimodal LLMs for GUI action prediction. Using a curated set of 136 challenging mobile-device tasks spanning five action types, UI-R1-3B reports average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL over the Qwen2.5-VL-3B base model, while remaining competitive with larger models trained via SFT on 76k samples. An efficiency-optimized variant UI-R1-E-3B is also introduced, with code released for reproducibility.

Significance. If the gains are attributable to the rule-based action reward and GRPO rather than data selection, the result is significant: it demonstrates that small-scale, high-quality RL can deliver substantial improvements in GUI agent tasks, offering a data-efficient alternative to large-scale SFT. The public code release supports direct verification of the reported benchmark numbers.

major comments (2)
  1. [Experiments] Experiments section: the manuscript provides no SFT ablation on the identical 136-task dataset using the same base model. Without this control, the reported lifts cannot be unambiguously attributed to the rule-based reward and GRPO rather than to the curation of high-quality examples; this directly undermines the central claim that rule-based RL is the source of the 22.1%/6.0%/12.7% gains.
  2. [Method] Method / Reward Design: the precise formulation of the rule-based action reward (including per-action-type scoring rules, any thresholds, and handling of partial matches) is not stated with sufficient formality to permit assessment of bias or reward hacking across the five action types.
minor comments (2)
  1. [Experiments] Table 1 or equivalent: add a row or column showing the exact number of examples per action type in the 136-task set to clarify balance.
  2. [Abstract] The abstract's contrast with 76k-sample SFT models is useful, but the text should explicitly note that no SFT baseline on the 136 tasks was run.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to strengthen the presentation of our results and methods.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the manuscript provides no SFT ablation on the identical 136-task dataset using the same base model. Without this control, the reported lifts cannot be unambiguously attributed to the rule-based reward and GRPO rather than to the curation of high-quality examples; this directly undermines the central claim that rule-based RL is the source of the 22.1%/6.0%/12.7% gains.

    Authors: We agree that an SFT ablation using the identical 136-task dataset and the same Qwen2.5-VL-3B base model would provide a stronger control experiment. Our current results demonstrate gains over the base model and competitiveness with larger SFT models trained on 76k samples, but this additional baseline will help isolate the contribution of the rule-based reward and GRPO more clearly. We will add this SFT ablation to the revised Experiments section. revision: yes

  2. Referee: [Method] Method / Reward Design: the precise formulation of the rule-based action reward (including per-action-type scoring rules, any thresholds, and handling of partial matches) is not stated with sufficient formality to permit assessment of bias or reward hacking across the five action types.

    Authors: We acknowledge that the reward formulation was presented at a descriptive level rather than with full formality. In the revised manuscript, we will add a precise mathematical definition of the rule-based action reward, explicitly detailing the per-action-type scoring rules for the five action types, any thresholds used, and the treatment of partial matches. This will facilitate assessment of potential biases or reward hacking. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper is an empirical study that curates a 136-task dataset, applies rule-based RL with GRPO and a custom action reward to Qwen2.5-VL-3B, and reports accuracy gains on independent external benchmarks (ScreenSpot, ScreenSpot-Pro, ANDROIDCONTROL). No equations, derivations, or self-citations are present that reduce the reported performance lifts to fitted parameters defined by the same data or to any self-referential construction. The results are measured on held-out evaluation sets whose labels are not used in training or reward computation, making the outcome chain self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the transferability of language-model RL techniques to multimodal GUI settings and on the sufficiency of a small curated dataset; few explicit free parameters beyond standard RL hyperparameters.

free parameters (1)
  • GRPO hyperparameters
    Standard policy optimization parameters tuned during training; exact values not stated in abstract.
axioms (1)
  • domain assumption Rule-based rewards derived from action correctness are sufficient to guide policy improvement in multimodal GUI tasks
    Transferred from language-model success without additional justification in the abstract.
invented entities (1)
  • Rule-based action reward no independent evidence
    purpose: Provide scalar feedback for GUI action predictions without human preference data
    Newly defined for this framework; no independent evidence outside the reported experiments.

pith-pipeline@v0.9.0 · 5647 in / 1269 out tokens · 76915 ms · 2026-05-16T10:58:22.234334+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 18 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Beyond Binary: Reframing GUI Critique as Continuous Semantic Alignment

    cs.LG 2026-05 unverdicted novelty 7.0

    BBCritic uses contrastive learning to align GUI actions in a continuous affordance space, outperforming larger binary critic models on a new four-level hierarchical benchmark while enabling zero-shot transfer.

  2. OS-SPEAR: A Toolkit for the Safety, Performance,Efficiency, and Robustness Analysis of OS Agents

    cs.CL 2026-04 unverdicted novelty 7.0

    OS-SPEAR is a new evaluation toolkit that tests 22 OS agents and identifies trade-offs between efficiency and safety or robustness.

  3. ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence on Mobile Devices

    cs.AI 2026-02 conditional novelty 7.0

    ProactiveMobile is a new benchmark for proactive mobile agents that tests latent intent inference from context and executable API generation, where a fine-tuned 7B model reaches 19.15% success versus 15.71% for o1 and...

  4. GUI-R1 : A Generalist R1-Style Vision-Language Action Model For GUI Agents

    cs.CV 2025-04 unverdicted novelty 7.0

    GUI-R1 uses reinforcement fine-tuning with GRPO on a small curated dataset to create a generalist vision-language action model that outperforms prior GUI agent methods across mobile, desktop, and web benchmarks using ...

  5. Video-R1: Reinforcing Video Reasoning in MLLMs

    cs.CV 2025-03 conditional novelty 7.0

    Video-R1 uses temporal-aware RL and mixed datasets to boost video reasoning in MLLMs, with a 7B model reaching 37.1% on VSI-Bench and surpassing GPT-4o.

  6. ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents

    cs.AI 2026-05 unverdicted novelty 6.0

    ToolCUA introduces a trajectory scaling pipeline and staged RL to optimize GUI-tool switching, reaching 46.85% accuracy on OSWorld-MCP for a 66% relative gain over baseline.

  7. LiteGUI: Distilling Compact GUI Agents with Reinforcement Learning

    cs.AI 2026-05 unverdicted novelty 6.0

    LiteGUI trains 2B/3B-scale GUI agents via SFT-free guided on-policy distillation and multi-solution dual-level GRPO to reach SOTA lightweight performance and compete with larger models.

  8. BAMI: Training-Free Bias Mitigation in GUI Grounding

    cs.CV 2026-05 unverdicted novelty 6.0

    BAMI mitigates precision and ambiguity biases in GUI grounding via coarse-to-fine focus and candidate selection, raising accuracy on ScreenSpot-Pro without training.

  9. ROSE: Rollout On Serving GPUs via Cooperative Elasticity for Agentic RL

    cs.DC 2026-05 unverdicted novelty 6.0

    ROSE delivers 1.2-3.3x higher end-to-end throughput for agentic RL by safely co-using underutilized serving GPUs for rollouts while meeting serving SLOs.

  10. AutoGUI-v2: A Comprehensive Multi-Modal GUI Functionality Understanding Benchmark

    cs.CV 2026-04 unverdicted novelty 6.0

    AutoGUI-v2 is a new benchmark exposing that VLMs handle basic GUI grounding but struggle with complex interaction logic and state prediction.

  11. UI-Zoomer: Uncertainty-Driven Adaptive Zoom-In for GUI Grounding

    cs.CV 2026-04 unverdicted novelty 6.0

    UI-Zoomer uses uncertainty quantification to trigger and size adaptive zoom-ins only on uncertain GUI grounding predictions, yielding up to 13.4% gains on benchmarks with no training.

  12. On Training Large Language Models for Long-Horizon Tasks: An Empirical Study of Horizon Length

    cs.AI 2026-05 unverdicted novelty 5.0

    Longer action horizons bottleneck LLM agent training through instability, but training with reduced horizons stabilizes learning and enables better generalization to longer horizons.

  13. Measure Twice, Click Once: Co-evolving Proposer and Visual Critic via Reinforcement Learning for GUI Grounding

    cs.LG 2026-04 unverdicted novelty 5.0

    A co-evolving proposer-critic RL framework improves GUI grounding accuracy by letting the model critique its own proposals rendered on screenshots.

  14. HalluClear: Diagnosing, Evaluating and Mitigating Hallucinations in GUI Agents

    cs.AI 2026-04 unverdicted novelty 5.0

    HalluClear supplies a taxonomy, calibrated evaluation, and lightweight post-training mitigation that reduces hallucinations in GUI agents using only 9K samples.

  15. From Pixels to Digital Agents: An Empirical Study on the Taxonomy and Technological Trends of Reinforcement Learning Environments

    cs.AI 2026-03 unverdicted novelty 5.0

    An empirical literature analysis reveals a bifurcation in RL environments into Semantic Prior (LLM-dominated) and Domain-Specific Generalization ecosystems with distinct cognitive fingerprints.

  16. Securing Computer-Use Agents: A Unified Architecture-Lifecycle Framework for Deployment-Grounded Reliability

    cs.CL 2026-05 unverdicted novelty 4.0

    The paper develops a unified framework that organizes computer-use agent reliability around perception-decision-execution layers and creation-deployment-operation-maintenance stages to map security and alignment inter...

  17. A Brief Overview: Agentic Reinforcement Learning In Large Language Models

    cs.AI 2026-04 unverdicted novelty 2.0

    The paper surveys the conceptual foundations, methodological innovations, challenges, and future directions of agentic reinforcement learning frameworks that embed cognitive capabilities like meta-reasoning and self-r...

  18. A Brief Overview: Agentic Reinforcement Learning In Large Language Models

    cs.AI 2026-04 unverdicted novelty 2.0

    This review synthesizes conceptual foundations, methods, challenges, and future directions for agentic reinforcement learning in large language models.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 17 Pith papers · 8 internal anchors

  1. [1]

    L1: Controlling how long a reasoning model thinks with reinforcement learning.arXiv preprint arXiv:2503.04697, 2025

    Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697,

    work page arXiv

  2. [2]

    Amex: Android multi-annotation expo dataset for mobile gui agents

    Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490,

    work page arXiv

  3. [3]

    R1-v: Reinforcing super gen- eralization ability in vision-language models with less than $3

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super gen- eralization ability in vision-language models with less than $3. https://github.com/ Deep-Agent/R1-V, 2025a. Accessed: 2025-02-02. Zhangquan Chen, Xufang Luo, and Dongsheng Li. Visrl: Intention-driven visual perception via reinforced reasoning. arXiv preprint arXiv:2503.0...

    work page arXiv 2025

  4. [4]

    Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents

    11 Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243,

    work page internal anchor Pith review arXiv

  5. [5]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Screenspot-pro: Gui grounding for professional high-resolution computer use, 2025a

    Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. Screenspot-pro: Gui grounding for professional high-resolution computer use, 2025a. Ming Li, Jike Zhong, Shitian Zhao, Yuxiang Lai, and Kaipeng Zhang. Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning. arX...

    work page arXiv

  8. [8]

    MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-based Reinforcement Learning

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, et al. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning. arXiv preprint arXiv:2503.07365,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    s1: Simple test-time scaling

    URL https://arxiv.org/abs/2501.19393. Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    UI-TARS: Pioneering Automated GUI Interaction with Native Agents

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Nils Reimers and Iryna Gurevych

    12 Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614,

    work page arXiv

  12. [12]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Vlm-r1: A stable and generalizable r1-style large vision-language model

    Haozhan Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. Vlm-r1: A stable and generalizable r1-style large vision-language model. https: //github.com/om-ai-lab/VLM-R1 , 2025a. Accessed: 2025-02-15. Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. Dast: Difficu...

    work page arXiv 2025

  15. [15]

    Wenyi Xiao, Leilei Gan, Weilong Dai, Wanggui He, Ziwei Huang, Haoyuan Li, Fangxun Shu, Zhelun Yu, Peng Zhang, Hao Jiang, et al

    URL https://arxiv.org/abs/2503.00401. Wenyi Xiao, Leilei Gan, Weilong Dai, Wanggui He, Ziwei Huang, Haoyuan Li, Fangxun Shu, Zhelun Yu, Peng Zhang, Hao Jiang, et al. Fast-slow thinking for large vision-language model reasoning. arXiv preprint arXiv:2504.18458,

    work page arXiv

  16. [16]

    Aguvis: Unified pure vi- sion agents for autonomous gui interaction.arXiv preprint arXiv:2412.04454, 2024

    Yiheng Xu, Zekun Wang, Junli Wang, Dunjie Lu, Tianbao Xie, Amrita Saha, Doyen Sahoo, Tao Yu, and Caiming Xiong. Aguvis: Unified pure vision agents for autonomous gui interaction. arXiv preprint arXiv:2412.04454,

    work page arXiv

  17. [17]

    AppAgent: Multimodal Agents as Smartphone Users

    Notion Blog. Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771,

    work page arXiv

  18. [18]

    aha moment

    13 Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1-zero’s” aha moment” in visual reasoning on a 2b non-sft model. arXiv preprint arXiv:2503.05132,

    work page arXiv