Source-linked AI summary
Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents
AIMAE Team, Tianxiang Chen, Yan Cheng, Zhangye Han, Xiaowei Li, Chang Liu, Cheng Liu, Zhongqiang Ma, Long Peng, Xiaobing Tu, Yinggui Wang, Hongliang Wei, Chen Wu, Daiping Xin, Kunyu Zhou, Pengyang Zhou, Peiyuan Chen, Ziyuan Chen, Yutao Deng, Chunyu Dong, Xiangyu Fu, Yicheng Feng, Ruian He, Haochen Li, Miancan Liu, Zhengqin Liu, Wei Peng, Jinkui Ren, Haoyu Tan, Dong Xiao, Rongkun Xue, Shujian Yang, Xianhang Ye, Ziqi Yuan, Ziyang Yu, Linghan Zhang, Xiantao Zhang, Xuanpu Zhao, Yinan Zhao, Zhenghui Zhao, Bin Zhu, Likai Zou
TL;DR
Real-world browser agents must handle long, changing interactions, yet training and evaluation remain centered on short successful demonstrations. Wuying-Browser-Agent co-designs execution, supervision, optimization, and evaluation, achieving a 70.8% average success rate versus 70.3% for the strongest competing open model.
Problem
Browser-agent training and benchmarks remain concentrated on short, successful, English-centric interactions, despite real deployment requiring long-horizon decisions, recovery, and complex UI navigation.
Method
Wuying-Browser-Agent co-designs a structured browser harness, RUIC-SFT recovery and UI curriculum supervision, DAO-GRPO optimization, and bilingual long-horizon evaluation.
Results
70.8% average success rate makes Wuying-Browser-Agent-27B the strongest open-source model across three browser-use benchmarks, exceeding Qwen3.8-Max at 70.3%.
Takeaways & Limitations
The pipeline provides a unified approach for improving long-horizon browser-agent robustness through recovery supervision, behavior-focused optimization, and deployment-aligned evaluation.
Takeaways & Limitations
Evaluation controls execution infrastructure by using the same harness, tool space, and observation pipeline across compared models.
Abstract
from arXiv · showhide
Browser agents perform well on short, clean demonstrations, but real deployment is fundamentally different: agents must sustain dozens of decisions on live websites while recovering from mistakes and navigating complex UIs. We argue that closing this gap requires alignment at every level of the pipeline, including execution, supervision, optimization, and evaluation, rather than scale alone. We present Wuying-Browser-Agent, a unified framework that addresses each of these levels. A structured browser harness provides stable execution primitives and decision-oriented context management. Reflection and UI-specialized Curriculum SFT (RUIC-SFT) explicitly trains on recovery trajectories and complex-UI interactions. Divergence-Aware Online GRPO (DAO-GRPO) improves long-horizon credit assignment through potential-based reward shaping and divergence-aware step weighting. Finally, we introduce BrowserBench, a bilingual real-web benchmark of 350 tasks averaging 37.9 steps, because most existing benchmarks are too short to expose long-horizon failure modes. Wuying-Browser-Agent-27B achieves 80.6\% on WebVoyager, 66.7\% on Online-Mind2Web, and 65.1\% on BrowserBench, establishing a new open-source state of the art on browser-use benchmarks. The same pipeline also transfers beyond browser use, demonstrating strong general agentic ability and reaching an average score of 73.8 on Tau2-Bench, Claw-Eval, and BFCL-v4.
1 Introduction
Real-world browser deployment requires agents to sustain long, changing, multilingual web interactions while recovering from errors and focusing learning on decisive actions. Wuying-Browser-Agent addresses these gaps through a co-designed execution, supervision, optimization, and evaluation pipeline.
- Results: Experiments show consistent gains across WebVoyager, Online-Mind2Web, and BrowserBench, with RUIC-SFT improving recovery and complex-UI performance and DAO-GRPO adding gains as difficulty and trajectory length increase.BrowserBench and qualitative analyses further reveal capability differences obscured by aggregate scores alone.
- Challenges: Long-horizon deployment exposes three gaps: insufficient recovery supervision, diffuse final-outcome feedback, and underrepresentative English-centric evaluation.These gaps arise because real tasks involve changing pages, unexpected redirects, complex widgets, and behaviorally decisive branch steps.
- Unified pipeline: The structured browser harness provides a validated action interface and decision-oriented state management across training, online reinforcement learning, and evaluation.It also maintains efficient execution contexts for supervised data construction, online rollouts, and evaluation.
- Unified pipeline: RUIC-SFT combines reflection-rich recovery supervision with UI-specialized interaction data under a progressive curriculum mixing schedule.The method is designed to address successful-trajectory training data that provides little guidance after execution errors.
- Unified pipeline: DAO-GRPO is an online reinforcement-learning framework for sparse-feedback, long-horizon browser tasks that concentrates learning on behaviorally decisive steps.Its motivation is that uniform trajectory-level supervision spreads learning across long shared segments instead of the decisions determining success or failure.
- Evaluation: BrowserBench introduces 350 bilingual Chinese–English real-web tasks averaging 37.9 steps to evaluate realistic long-horizon interactions beyond English-centric benchmarks.The suite targets the underrepresentation of diverse real websites and longer, compositional workflows in existing evaluations.
2 Related Work
Related work spans foundation-model browser agents, native GUI and computer-use systems, and outcome-based reinforcement learning, but existing approaches often fall short on recovery, long-horizon credit assignment, and bilingual real-web evaluation. BrowserBench addresses key evaluation gaps with 350 bilingual real-web tasks averaging 37.9 steps across 254 websites and structured success criteria.
- Browser Agents: Modern browser agents combine general-purpose or vision-language foundation models with reasoning-and-acting frameworks and web-specific grounding or post-training.ReAct established the reasoning-and-acting paradigm, while SeeAct demonstrated GPT-4V-based web agency with structured grounding.
- GUI and Computer-Use Agents: Native GUI and computer-use agents extend browser interaction across desktop, web, and mobile environments, using screenshot perception, large-scale supervision, or unified vision-language control.Examples include UI-TARS, OpenCUA, and GUI-Owl-1.5, alongside compact models trained with efficient designs or curated demonstrations.
- Research Gap: Most computer-use systems prioritize atomic control or general ability rather than jointly handling recovery-oriented supervision, sparse-reward credit assignment, and bilingual long-horizon evaluation.The paper positions Wuying-Browser-Agent as complementary to this line of work by targeting these three challenges in real-web execution.
- Reinforcement Learning: Outcome-based reinforcement learning has expanded from language reasoning to multimodal visual and GUI tasks, with GRPO and DAPO enabling critic-free group-relative policy optimization.Browser-specific work increasingly studies online RL because fixed demonstrations do not reliably teach exploration, recovery, and adaptation to changing pages.
- Browser Benchmarks: 350 bilingual real-web tasks average 37.9 steps and span 254 websites in BrowserBench, with goal-only instructions, structured success criteria, and capability-level annotations.The benchmark is designed to address the underrepresentation of long-horizon tasks, English-centric evaluation, and narrow website coverage.
3 Preliminaries
Browser-agent execution is modeled as a partially observable sequential decision process in which structured observations, optional screenshots, retained history, and feedback support action selection over bounded trajectories. A structured harness executes parsed browser actions, while an external judge evaluates whether the terminal state satisfies the task specification.
- Execution Model: At each of at most Tmax steps, the environment returns an observation, the agent emits a response, the controller parses a structured action, and feedback drives the next state.This defines the interaction loop as a sequential decision process.
- Observations and Context: The primary observation is a structured DOM-derived browser state, augmented with a viewport screenshot when visual layout, rendered content, or image cues are needed.The observation is therefore either St alone or (St, Vt) when visual grounding is required.
- Observations and Context: The policy conditions on a reconstructed decision context containing the task instruction, structured browser state, optional screenshot, prior actions, and environment feedback.Context reconstruction preserves information relevant to the current decision while avoiding redundant append-only history.
- Action Space: The unified action space covers navigation, interaction, extraction, file operations, and flow control, with actions grounded in the current observation and deterministically parsed by the harness.Table 1 specifies 24 atomic operations across these five categories.
- Trajectory and Objective: Task success requires a trajectory ending in a browser state that satisfies the task specification, as judged externally from the final state and interaction log without direct reward observation during execution.The agent must infer progress from browser observations and environment feedback.
4 Method
Wuying-Browser-Agent addresses long-horizon browser robustness through a shared execution harness, robustness-oriented RUIC-SFT supervision, and online optimization designed around recovery and realistic decision contexts. Its method combines structured execution and context management with specialized UI, reflection, and progress-aware training signals.
- Browser harness: The browser harness parses, validates, and executes model actions while returning structured feedback, rejecting malformed calls with typed errors that support correction.It connects policy outputs to live pages and is shared by RUIC-SFT and DAO-GRPO.
- Browser harness: The harness maintains decision-oriented context and intermediate artifacts instead of exposing an ever-growing raw transcript, supporting cross-step memory and file or report delivery.Its context includes relevant task state, action history, and environment feedback at each step.
- RUIC-SFT: RUIC-SFT supplements general demonstrations with specialized UI and reflection data to teach complex-control interaction and recovery after off-trajectory errors.The reflection data provides supervision for deviations largely absent from successful demonstrations, while UI data repeatedly exposes difficult structured interaction patterns.
- RUIC-SFT: Dynamic dataset mixing begins general-data-dominant, increases specialized UI data, and introduces reflection data late to strengthen interaction skills and self-correction without destabilizing execution.The schedule is designed to stabilize basic browser operations before emphasizing complex controls and reflection.
- DAO-GRPO: DAO-GRPO uses potential-based progress shaping with terminal boundary Φ(s_Ti) = 0 and grouped parallel rollouts under updated policies to align optimization with changing browser-state contexts.The design addresses train–test context mismatch by optimizing responses under step-specific contexts reconstructed from page state, history, and optional visual evidence.
5 BrowserBench: A Bilingual Long-Horizon Real-Web Benchmark
BrowserBench is a bilingual, long-horizon real-web benchmark designed to evaluate sustained browser interaction beyond predominantly English, short-task settings. It combines broad task coverage with goal-only, single-rollout evaluation for realistic and interpretable assessment.
- Language and horizon: The benchmark includes 191 Chinese tasks (54.6%) and 159 English tasks (45.4%), with task lengths ranging from 15 to 100 interaction steps.Chinese tasks are longer on average than English tasks.
- Task coverage: Tasks cover e-commerce, rankings and comparisons, data collection, maps and travel, organization verification, academic search, general search, and news browsing.These categories represent common real-world browser-use patterns across Chinese and English websites.
- Construction and curation: Tasks are collected from authentic browsing scenarios, normalized into goal-only instructions with structured success criteria, and manually verified through repeated browser execution.Curation removes procedural leakage, ambiguous goals, unsupported entities, and stale page structures to make failures more informative.
- Benchmark scope: BrowserBench contains 350 bilingual real-web tasks spanning 254 websites, with an average completion length of 37.9 steps.The benchmark targets sustained multi-step browsing over realistic Chinese–English websites.
- Evaluation protocol: Evaluations use Pass@1 Success Rate under a single-rollout, goal-only protocol with a maximum budget of 100 interaction steps.Runs that do not satisfy the structured success criterion within the budget are counted as failures.
6 A Self-Reinforcing Data Flywheel
Wuying-Browser-Agent forms a self-reinforcing data flywheel in which deployment, evaluation, and training continuously generate supervision for later cycles. Its automated logging, triage, and correction reduce recurring annotation demands, while targeted human judgment and multi-cycle preference training help maintain alignment with the evolving web.
- Flywheel design: The harness and BrowserBench create a closed loop where each deployment round produces supervision for the next.This addresses live websites’ continuous change, which causes fixed demonstration corpora to depreciate over time.
- Flywheel stages: The flywheel unifies trajectory collection, automated triage, correction re-execution, data distillation, and training-mixture updates.Episodes from DAO-GRPO training, benchmark evaluation, and deployment share structured, training-ready logging with actions, feedback, and reconstructed contexts.
- Human–automation division: Automation handles logging, triage, correction re-execution, and reproducible scoring, while humans author reflections, annotate complex UIs, and audit borderline judgments.A routed LLM judge estimates rewards over thousands of online rollouts when throughput matters.
- Bootstrapping cycle: About 3,000 rollout trajectories yielded roughly 420 verified reflection examples and 130 UI-component scenes for the released models’ training mixture.An initial Base SFT model trained only on general demonstrations bootstrapped this cycle, while early BrowserBench diagnostics guided the online task pool.
- Sustained operation: Each additional cycle requires only sandbox compute and lightweight human auditing, supporting sustained multi-cycle preference training over accumulated contrastive pairs.The authors identify this operation as the primary mechanism for keeping Wuying-Browser-Agent aligned with the evolving web.
7 Experiments · 7.1 Training Settings · 7.2 Evaluation Benchmarks
The experiments align training with long browser trajectories through RUIC-SFT and DAO-GRPO, then evaluate browser-use and general agentic capability under a shared execution harness. The benchmark suite includes real-web tasks of varying scale and length, alongside tests of transfer beyond web interaction.
- 7.1 Training Settings: RUIC-SFT fine-tunes Qwen3.5-series browser policies for two epochs with LoRA rank 16, α = 64, cosine scheduling, and 10% linear warmup.All linear layers are targeted, with sequence parallelism enabled on a multi-GPU cluster for long browser trajectories.
- 7.1 Training Settings: DAO-GRPO further optimizes the policy using parameter-efficient LoRA adaptation with rank 8 and α = 16.All linear layers are used as target modules.
- 7.1 Training Settings: Online training uses an asynchronous rollout scheduler that dispatches tasks in batches across isolated AgentBay sandboxes and streams completed rollouts to the optimizer.Sandboxes are recycled after each episode to guarantee state isolation.
- 7.1 Training Settings: The online reward combines format-validity checks with an LLM-as-a-judge success signal, while routed evaluation controls judging cost and abnormal or unreachable trajectories are masked out.The supplied passage ends after describing masking, so no further masking condition is reported.
- 7.2 Evaluation Benchmarks: All evaluated models use the same harness, tool space, and observation pipeline, making performance differences reflect policy capability rather than execution infrastructure.The experiments cover browser-use capability and general agentic capability.
- 7.2 Evaluation Benchmarks: Transfer beyond web interaction is assessed with Tau2-Bench for conversational tool use, BFCL-V4 for function-calling reliability, and Claw-Eval for end-to-end autonomous performance.Together, these benchmarks provide a complementary view of transfer, although the supplied passage truncates its concluding clause.
7.3 Compared Baselines
Wuying-Browser-Agent is evaluated against closed- and open-source browser-use baselines, as well as browser-agent and general-purpose models on broader agentic benchmarks.
- Compared Baselines: The evaluation compares Wuying-Browser-Agent against baselines reported in Tables 2 and 3.These comparisons cover both browser-use and general agentic benchmarks.
- Browser-use baselines: Browser-use comparisons include closed-source models such as GPT-4o, GPT-5, GPT-5.5, Ernie-5.0, and Seed2.1-Pro.The closed-source category follows Table 2.
- Browser-use baselines: Browser-use comparisons also include open-source models including OpenWebRL-8B, Kimi k2.5, Qwen3-VL variants, Hy3-295B-A21B, Qwen3.5 variants, Qwen3.8-Max, and DeepSeek-V4-Flash-0731.The listed open-source baselines span multiple model families and sizes.
- General agentic baselines: General agentic comparisons include browser-agent baselines OpenCUA 72B, GUI-Owl-1.5-32B-Instruct, UI-Venus-1.5-30B-A3B, and EvoCUA-32B-20260105.These models are reported in Table 3.
- General agentic baselines: General-purpose comparisons include Qwen3.5-27B and Qwen-UI-Agent-27B to assess transfer beyond browser interaction to broader tool use and autonomy.The comparison tests whether browser-grounded training transfers to general agentic ability.
7.4 Browser-Use Benchmark Results
Wuying-Browser-Agent-27B sets a new open-source state of the art across three complementary browser-use benchmarks, remaining competitive with strong closed-source systems. Online RL provides consistent gains at both 9B and 27B scales beyond supervised initialization.
- Open-source benchmark results: 70.8% average success rate makes Wuying-Browser-Agent-27B the best open-source model, ahead of Qwen3.8-Max at 70.3% and Qwen3.5-397B-A17B at 56.5%.It reaches 80.6% on WebVoyager, 66.7% on Online-Mind2Web, and 65.1% on BrowserBench.
- Benchmark coverage: The three benchmarks are complementary, covering English-centric live-web tasks alongside Chinese-web and deployment-oriented real-world websites.Strong performance across all three indicates generalization across benchmark conventions and web environments with different linguistic and interaction characteristics.
- Comparison with closed-source systems: 75.7% for GPT-5.5, 69.6% for GPT-5, and 64.4% for Qwen3.7-Plus position the open-source model competitively with strong closed-source systems.Wuying-Browser-Agent-27B surpasses Qwen3.7-Plus at 70.8% versus 64.4%.
- Scaling and online RL: 50.8% at 9B and 70.8% at 27B are the final-model results after online RL, improving over SFT from 45.6% and 62.7%, respectively.The consistent gains across scales indicate DAO-GRPO improves beyond supervised initialization, while both SFT and RL variants benefit from increased model size.
7.5 Transfer to General Agentic Benchmarks
Wuying-Browser-Agent’s browser-grounded training transfers beyond browser use to general agentic benchmarks. It remains competitive with strong general-purpose tool agents and improves over the Qwen3.5-27B base model across all reported benchmarks.
- Transfer Evaluation: The evaluation tests transfer beyond browser use across Tau2-Bench, Claw-Eval, and BFCL-v4.These benchmarks represent more general agentic settings.
- Transfer Evaluation: Wuying-Browser-Agent-27B remains competitive with strong general-purpose tool agents across the reported benchmarks.The results do not indicate over-specialization to web interaction.
- Transfer Evaluation: Wuying-Browser-Agent-27B improves over the Qwen3.5-27B base model on all reported benchmarks.The passage attributes these gains to harness-grounded supervision and long-horizon browse training.
7.6 Ablation Study
The ablations show that curriculum scheduling and complementary optimization components improve browser-agent success, recovery, efficiency, and performance on difficult tasks. DAO-GRPO further strengthens long-horizon learning through dense progress signals, localized credit assignment, and response-level optimization.
- RUIC-SFT ablations: 29.3 average actions under fixed global mixing versus 24.8 for single-source training suggests uniform mixtures introduce redundant or hesitant execution.Fixed mixing improves overall completion but increases action count.
- RUIC-SFT ablations: 38.0% overall success and 18.5% recovery success make full RUIC-SFT the strongest schedule, while average steps fall to 22.4.The curriculum outperforms fixed-ratio multi-source training by balancing execution efficiency with recovery behavior.
- DAO-GRPO ablations: 40.7% overall success with PBRS improves on vanilla online GRPO’s 39.4%, showing that dense progress signals help when terminal rewards are sparse.Vanilla online GRPO improves only modestly over the 38.0% RUIC-SFT initialization.
- DAO-GRPO ablations: 27.6% hard-task and 29.7% recovery-task success after divergence-aware credit assignment show its largest gains on targeted subsets.The method concentrates credit around decisive branch points rather than weighting all response segments uniformly.
- DAO-GRPO ablations: 42.9% overall, 30.5% hard-task, and 32.8% recovery-task success under reconstructed-context response optimization establish the strongest full objective.Optimizing each response under its own decision context improves all reported metrics.
- Difficulty ablations: +12.4 points on hard tasks is DAO-GRPO’s largest difficulty-slice improvement, raising performance from 18.1% to 30.5%.Its gains are +1.0 point on easy tasks and +2.1 points on medium tasks, indicating stronger benefits for long-horizon correction and recovery.
8 Conclusion
The paper identifies the mismatch between success-dominated short-horizon training and real long-horizon browser deployment as the central robustness bottleneck. It addresses this through a unified pipeline co-designing supervision, online optimization, and evaluation for long-horizon deployment.
- Conclusion: The central bottleneck is the mismatch between success-dominated short-horizon training and real long-horizon browser deployment.The paper links this mismatch to limited real-world robustness in web agents.
- Conclusion: The proposed unified pipeline co-designs supervision, online optimization, and evaluation for the same long-horizon deployment setting.The pipeline is intended to address the training–deployment mismatch at multiple levels.
- Conclusion: RUIC-SFT provides a capability-structured initialization at the supervision level.The supplied passage states that RUIC-SFT integrates reflection-rich training, but the excerpt ends before specifying further details.
Contributions and Acknowledgments
The acknowledgments list contributors alphabetically by last name. The roster spans AIMAE Team and named contributors across three entries.
- Acknowledgments: Contributors are listed in alphabetical order by their last names.This ordering applies to the acknowledgment roster.
- Acknowledgments: The first roster entry includes AIMAE Team, Tianxiang Chen, Yan Cheng, Zhangye Han, Xiaowei Li, Chang Liu, Cheng Liu, Zhongqiang Ma, Long Peng, and Xiaobing Tu†.The entry presents ten contributors or contributor groups.
- Acknowledgments: The second roster entry lists Yinggui Wang, Hongliang Wei, Chen Wu, Daiping Xin, Kunyu Zhou, and Pengyang Zhou.These names continue the contributor acknowledgment list.
- Acknowledgments: The final roster entry lists Peiyuan Chen, Ziyuan Chen, Yutao Deng, Chunyu Dong, Xiangyu Fu, Yicheng Feng, Ruian He, Haochen Li, Miancan Liu, Zhengqin Liu, Wei Peng, Jinkui Ren, Haoyu Tan, Dong Xiao, Rongkun Xue, Shujian Yang, Xianhang Ye, Ziqi Yuan, Ziyang Yu, Linghan Zhang, Xiantao Zhang, Xuanpu Zhao, Yinan Zhao, Zhenghui Zhao, Bin Zhu, and Likai Zou.This entry contains the remaining named contributors.