PaperlayerView exact revision on arXiv

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

arXiv:2608.17319v1cs.AI

TL;DR

Real-world browser tasks require long sequences of decisions and recovery on live sites, unlike the short, successful interactions emphasized by existing training and benchmarks. Wuying-Browser-Agent co-designs a long-horizon pipeline and establishes a new open-source state of the art across browser-use benchmarks.

  • Problem

    Existing browser-agent training and benchmarks center on short, successful, English-centric interactions, while real deployment requires sustained decisions and recovery on live, complex websites.

    Source: §1 · p. 2

  • Method

    The paper co-designs a long-horizon browser-agent pipeline around a shared harness, recovery- and UI-specialized curriculum supervision, and divergence-aware online optimization.

    Source: §1 · p. 3, §8 · p. 27

  • Results

    Across three browser-use benchmarks, Wuying-Browser-Agent-27B establishes a new open-source state of the art, averaging 70.8% success versus Qwen3.8-Max's 70.3%.

    Source: §1 · p. 3, §7.4 · p. 18

  • Takeaways & Limitations

    BrowserBench exposes capability differences that aggregate scores alone obscure.

    Source: §1 · p. 3

    Evaluation comparisons use a shared harness, tool space, and observation pipeline, so reported performance differences reflect policy capability rather than execution infrastructure.

    Source: §7.2 · p. 17

Abstract

from arXiv · show

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 tasks require agents to sustain many decisions on changing websites, yet training and evaluation remain concentrated on short, successful, English-centric interactions. Wuying-Browser-Agent co-designs execution, supervision, optimization, and evaluation for this long-horizon setting.

Results

  • At 27B, Wuying-Browser-Agent establishes a new open-source state of the art on browser-use benchmarks and transfers to broader tool-use benchmarks.The transfer benchmarks include Tau2-Bench, BFCL-v4, and Claw-Eval.Source: §1 · pp. 3–4

Challenges

  • Long-horizon deployment exposes three linked gaps: little supervision for recovery and complex UIs, sparse outcome feedback that obscures decisive steps, and benchmarks lacking bilingual real-web tasks.A base supervised model recovers from only 8.5% of detected erroneous steps; BrowserBench contains 350 tasks averaging 37.9 steps.Source: §1 · pp. 2–3

Structured browser harness

  • A structured browser harness provides a validated action interface and decision-oriented state management shared across supervised training, online rollouts, and evaluation.It uses compact updates for local page changes and avoids reasoning over obsolete DOM snapshots after navigation.Source: §1 · p. 3

RUIC-SFT

  • RUIC-SFT combines reflection-rich recovery trajectories with complex-UI interaction data in a progressive curriculum for basic operations, UI interaction, and self-correction.Its curriculum stabilizes basic operations before strengthening UI interaction and, finally, self-correction.Source: §1 · p. 3

DAO-GRPO

  • DAO-GRPO improves long-horizon optimization under sparse feedback with denser progress supervision, greater weight on behaviorally critical decisions, and step-wise decision contexts.Its gains grow with task difficulty and trajectory length.Source: §1 · p. 3

BrowserBench

  • BrowserBench evaluates long-horizon tasks on real Chinese and English websites with goal-only instructions, addressing the shortage of realistic bilingual online benchmarks.It comprises 350 Chinese–English tasks averaging 37.9 steps.Source: §1 · pp. 2–4

2 Related Work

Related work spans browser agents, computer-use systems, agentic reinforcement learning, and browser benchmarks, while identifying gaps in recovery, sustained online optimization, and bilingual long-horizon real-web evaluation.

2.1 Browser Agents

  • Browser agents build on general LLM/VLM backbones [41] [44] [3] [36], reasoning-and-acting [46], and structured grounding [50], while Wuying jointly addresses recovery data, online optimization, and evaluation.The framework combines BrowserBench, RUIC-SFT, and DAO-GRPO to address these gaps.Source: §2.1 · p. 4

2.2 Computer-Use and Generalist Browser Agents

  • Computer-use agents broaden GUI control across platforms [37] [42] [32] [14] [51], but often target atomic control or short-horizon, single-platform evaluation rather than sustained real-web recovery.Wuying complements this work with a validated browser harness while retaining competitive general agentic capability.Source: §2.2 · p. 4

2.3 Agentic Reinforcement Learning

  • Browser RL builds on SFT plus GRPO-style training [28] [26] [49] [40] [38], while DAO-GRPO adds PBRS shaping, divergence-based branch-sensitive credit assignment, and response-level objectives for changing browser contexts.Browser tasks are stateful and path-dependent, making fixed demonstrations alone insufficient for long-horizon exploration, recovery, and adaptation.Source: §2.3 · p. 5

2.4 Browser Agent Benchmarks

  • BrowserBench addresses underrepresented long-horizon, English-centric, narrow-site evaluation with 350 bilingual real-web tasks averaging 37.9 steps across 254 websites.Tasks use goal-only instructions with structured success criteria for automated scoring and are annotated by difficulty and task pattern.Source: §2.4 · p. 5

3 Preliminaries

The preliminaries frame browser use as a finite-horizon interaction process in which policies rely on structured observations and decision context. They also describe the browser environment interface and a shared structured action vocabulary.

3.1 Problem Formulation

  • At each of at most Tmax steps, the environment provides an observation, the agent response is parsed into an action, and feedback drives the next browser-state transition.Source: §3.1 · p. 6
  • Policy context combines the task instruction, structured browser state, optional screenshot for visual grounding, and retained actions and feedback to keep dynamic history decision-relevant.Source: §3.1 · p. 6
  • The agent uses a unified structured action space for navigation, interaction, extraction, file operations, and flow control, with actions grounded in observations and parsed deterministically.Source: §3.1 · p. 6
  • Task success requires a terminal browser state satisfying the specification, judged externally from the final state and interaction log, while the agent infers progress without direct reward.Source: §3.1 · p. 6

3.2 Browser Environment Interface

  • The lightweight MCP interface couples action execution with state-transition retrieval, avoiding repeated full-state transmission and reducing rollout communication overhead.Source: §3.2 · p. 6
  • The shared action vocabulary contains 24 atomic operations organized into five categories: navigation, interaction, extraction, file operations, and flow control.Source: Table 1 · p. 7

4 Method

Wuying-Browser-Agent trains a shared browser policy through robustness-oriented RUIC-SFT followed by DAO-GRPO, with a structured harness supporting execution, learning, and evaluation. RUIC-SFT teaches recovery and complex-UI interaction, while DAO-GRPO focuses online updates on decisive steps under reconstructed browser contexts.

4.1 Browser Harness Layer

  • The shared harness validates and executes structured browser calls, returns typed errors and updated observations, and manages task state and artifacts across training and evaluation.Its decision-oriented context avoids unbounded raw transcripts and gives the policy a consistent interface across offline learning, online interaction, and deployment.Source: §4.1 · p. 7, §4.1.2 · p. 8, §4.1.3 · p. 8

4.2.2 Reflection Dataset

  • RUIC-SFT adds reflection and specialized-UI supervision because successful demonstrations alone offer little recovery guidance and sparse, skewed coverage of complex controls.Reflection examples pair an erroneous action and resulting observation with grounded reasoning and a verified corrective continuation, teaching the policy to recognize off-path execution.Source: §4.2 · p. 8, §4.2.1 · p. 9, §4.2.2 · p. 9

4.2.3 Specialized UI Component Dataset

  • The specialized UI dataset builds reusable interaction patterns by collecting and verifying real control scenes, annotating increasingly difficult tasks, and augmenting examples across sites and anomalous states.It targets controls such as date pickers, cascaders, and rich-text editors, reducing dependence on sparse natural exposure.Source: §4.2.3 · pp. 9–10

4.2.4 Curriculum-Based Training Schedule

  • RUIC-SFT dynamically mixes general, specialized-UI, and reflection data, delaying reflection until browser execution skills stabilize; the objective remains standard supervised next-token prediction.This phased schedule aims to preserve the execution prior while adding complex-control skills and self-correction, and supplies DAO-GRPO with step-level recovery signals.Source: Figure 4 · p. 8, §4.2.4 · p. 10, §4.2.5 · pp. 10–11

4.3 Divergence-Aware Online GRPO (DAO-GRPO)

  • DAO-GRPO refines the RUIC-SFT initialization through live browser interaction, targeting branch-defining decisions under sparse feedback and changing contexts.Tasks are rolled out in parallel across browser sandboxes, with grouped trajectories consumed by the optimizer.Source: §4.3 · p. 11, §4.3.1 · p. 12

4.3.2 Grouped Return Estimation

  • DAO-GRPO combines terminal success with potential-based progress shaping and normalizes returns within same-task rollout groups to estimate relative trajectory advantages.The shaping is a practical approximation to policy-preserving PBRS in the finite-horizon grouped setting; groups with identical returns are discarded.Source: §4.3.2 · p. 12

4.3.3 Divergence-Aware Step Credit Assignment

  • DAO-GRPO uses semantic divergence estimates to emphasize branch-defining steps while preserving trajectory-level advantages as the direction of optimization.Shared-prefix steps receive less weight, divergence steps are weighted by estimated criticality, and favorable-branch steps retain credit longer; the estimator is a credit prior, not an oracle.Source: §4.3.3 · pp. 12–13

4.3.4 Response-Level Policy Optimization

  • DAO-GRPO evaluates each response using the exact step-specific context available at decision time, reducing the train–test mismatch of optimizing against a concatenated transcript.Contexts are reconstructed using harness rules, while group-relative advantages and step weights concentrate updates on decisive steps; a single on-policy update per rollout batch needs no importance correction.Source: §4.3.4 · pp. 13–14

5 BrowserBench: A Bilingual Long-Horizon Real-Web Benchmark

BrowserBench targets the underrepresented bilingual, long-horizon, real-web setting with 350 tasks across Chinese and English websites. Its curated tasks, taxonomy, difficulty calibration, and single-rollout protocol are designed to support realistic, interpretable evaluation.

5.1 Benchmark Construction and Curation

  • BrowserBench tasks come from authentic browsing scenarios and are normalized into goal-only instructions with structured success criteria, then manually reviewed and repeatedly executed.Curation removes excessive procedural leakage, ambiguous goals, unsupported entities, and stale page structures; retained cases are checked for executability, textual quality, and alignment with the browser environment.Source: §5.1 · p. 14

5 BrowserBench: A Bilingual Long-Horizon Real-Web Benchmark

  • 350 tasks span 254 websites, with 37.9 interaction steps on average, making BrowserBench a bilingual benchmark for sustained real-web browsing.It includes 191 Chinese tasks (54.6%) and 159 English tasks (45.4%); task lengths range from 15 to 100 steps, and Chinese tasks average 41.13 steps versus 34.08 for English tasks.Source: §5 · p. 14, Figure 6 · p. 15

5.2 Task Taxonomy and Dataset Statistics

  • 105 tasks are easy, 140 medium, and 105 hard, with difficulty calibrated by three reference agents and boundary-case human review.The 30% hard-task share exceeds Online-Mind2Web’s 26.3%; the split is balanced across languages, and average task length rises from easy to hard.Source: §5.2 · p. 15
  • The taxonomy covers eight common real-world browser task categories, enabling analysis across task structures with different planning, interaction, and extraction demands.The benchmark also reports category distributions and language settings to support interpretable slices beyond an aggregate score.Source: §5.2 · pp. 14–15

5.3 Evaluation Protocol and Metric

  • BrowserBench uses goal-only instructions and Pass@1 Success Rate, counting success only when a task’s structured criterion is met in one rollout within 100 steps.Tasks unfinished within the budget count as failures, reflecting the single-trajectory execution constraints of realistic browser deployment.Source: §5.3 · p. 16

6 A Self-Reinforcing Data Flywheel

The training pipeline, harness, and BrowserBench form a self-reinforcing loop that turns ongoing website interaction, including failures, into fresh, verified training data. Its triage and diagnostic feedback support adaptation as live websites change.

6 A Self-Reinforcing Data Flywheel

  • The flywheel links deployment, training, and evaluation in a closed loop that converts ongoing interaction, including failed episodes, into fresh, verified training signal.The paper argues that fixed demonstration corpora depreciate as live websites change, so sustained capability requires ongoing supervision from interaction.Source: §6 · p. 16
  • Four stages collect standardized trajectories, triage failures, verify candidate corrections before training, and use evaluation diagnostics to target future data collection.Triage routes environment-induced failures to task-pool maintenance, successes to general demonstrations, localizable errors to reflection, and widget-specific failures to UI scenes; only sandbox-reexecuted corrections that restore a viable path enter reflection data.Source: §6 · p. 16
  • Automation handles logging, triage, correction re-execution, and high-throughput reward estimation, while humans provide reflection reasoning, complex-UI annotations, and borderline audits.BrowserBench uses deterministic structured criteria for reproducible scoring, whereas a routed LLM judge estimates rewards across thousands of online rollouts.Source: §6 · pp. 16–17
  • About 3,000 bootstrapped rollout trajectories yielded 420 verified reflection examples and 130 UI-component scenes for released-model training mixtures.Diagnostic slices from an early BrowserBench snapshot also guided the online task pool’s composition.Source: §6 · p. 17
  • Each additional cycle requires sandbox compute and lightweight human auditing rather than proportional annotation effort, supporting sustained alignment with the evolving web.The authors identify sustained multi-cycle operation, including preference-based training over accumulated contrastive pairs, as the primary mechanism for maintaining alignment.Source: §6 · p. 17

7 Experiments · 7.1 Training Settings

The experiments initialize the browser policy with RUIC-SFT and further optimize it with DAO-GRPO using LoRA. Online training uses asynchronous isolated sandboxes, shaped rewards, and trajectory-level sampling, with a 50-step rollout cap distinct from the 100-step evaluation budget.

7.1 Training Settings

  • RUIC-SFT fine-tunes Qwen3.5-series models for two epochs with rank-16 LoRA, α = 64, all linear layers targeted, and a cosine schedule peaking at 5 × 10−5 after 10% warmup.Training uses a multi-GPU cluster with sequence parallelism to accommodate long browser trajectories.Source: §7.1 · p. 17
  • DAO-GRPO further optimizes the policy with rank-8 LoRA, α = 16, targeting all linear layers.Online training runs against a large-scale AgentBay sandbox pool.Source: §7.1 · p. 17
  • An asynchronous rollout scheduler dispatches tasks in batches, runs them in parallel across isolated AgentBay sandboxes, and streams completed rollouts to the optimizer.Sandboxes are recycled after each episode to guarantee state isolation.Source: §7.1 · p. 17
  • The reward combines format-validity checks with an LLM-as-a-judge success signal, while routed evaluation keeps judging costs tractable at scale.Trajectories that terminate abnormally or become unreachable have their advantages set to zero.Source: §7.1 · p. 17
  • Training rollouts are capped at 50 interaction steps, while evaluation uses the full 100-step budget.The training cap is a throughput–coverage trade-off, not a deployment-horizon limit; over-cap tasks are truncated as non-terminal, retaining accumulated shaping progress.Source: §7.1 · p. 17
  • Trajectory-level dynamic sampling uses the held-out validation set and discards groups with identical returns at each update.Source: §7.1 · p. 17

7.2 Evaluation Benchmarks

The section evaluates Wuying-Browser-Agent on online browser-use benchmarks and complementary general-agentic benchmarks, using a shared interaction pipeline to support fair comparison.

Evaluation setup

  • All evaluated models use the same harness, tool space, and observation pipeline, so performance differences reflect policy capability rather than execution infrastructure.Source: §7.2 · p. 17

Browser-use benchmarks

  • Browser-use evaluation covers WebVoyager, Online-Mind2Web, and BrowserBench, with the latter comprising 350 bilingual tasks across 254 websites averaging 37.9 steps.WebVoyager has 643 tasks across 15 websites, while Online-Mind2Web has 300 tasks across 136 websites. Unless specified otherwise, results use Pass@1 Success Rate under the official protocol; an audit found 96.4% agreement between judge decisions and human annotations, with no systematic bias favoring Qwen-family policies.Source: §7.2 · p. 17

General agentic benchmarks

  • Tau2-Bench, BFCL-V4, and Claw-Eval assess whether browser-grounded training transfers to broader conversational tool use, function calling, and autonomous agent tasks.Together, these benchmarks provide a complementary view of whether long-horizon browser improvements come without sacrificing broader agentic capability.Source: §7.2 · pp. 17–18

7.3 Compared Baselines

The section defines comparisons against browser-use baselines and general-agentic baselines, spanning closed- and open-source models, browser agents, and general-purpose agents.

7.3 Compared Baselines

  • Browser-use comparisons span closed-source and open-source models, including browser-specialized systems and general-purpose multimodal models at varied scales.Closed-source baselines are GPT-4o [15], GPT-5 [30], GPT-5.5, Ernie-5.0 [34], and Seed2.1-Pro. Open-source baselines are OpenWebRL-8B [45], Kimi k2.5 [31], Qwen3-VL-8B-Thinking [3], Qwen3-VL-235B-A22B-Thinking [3], Hy3-295B-A21B, Qwen3.5-4B/9B/27B/397B-A17B [44], Qwen3.8-Max, and DeepSeek-V4-Flash-0731 [41]; comparisons are listed in Tables 2 and 3.Source: §7.3 · p. 18
  • General-agentic comparisons assess whether browser-grounded training transfers to broader tool-use and autonomous-agent settings.Browser-agent baselines are OpenCUA 72B [37], GUI-Owl-1.5-32B-Instruct [42], UI-Venus-1.5-30B-A3B [32], and EvoCUA-32B-20260105 [14]; general-purpose agent baselines are Qwen3.5-27B [44] and Qwen-UI-Agent-27B [51].Source: §7.3 · p. 18

7.4 Browser-Use Benchmark Results

Wuying-Browser-Agent-27B establishes a new open-source state of the art across browser-use benchmarks, with strong results across English-centric and Chinese-web settings. Its performance compares favorably with open and closed-source systems, and online RL gains hold across model scales.

7.4 Browser-Use Benchmark Results

  • WebVoyager and Online-Mind2Web are English-centric live-web benchmarks, while BrowserBench emphasizes Chinese-web and deployment-oriented real-world websites.Strong results across all three indicate generalization across benchmark conventions, linguistic characteristics, and interaction environments.Source: §7.4 · p. 18
  • 70.8% versus 64.4% for Qwen3.7-Plus; Wuying-Browser-Agent-27B remains reasonably close to GPT-5 (69.6%) and GPT-5.5 (75.7%).The paper notes that Wuying-Browser-Agent is fully open-source and trained with a unified pipeline targeting long-horizon real-browser interaction.Source: §7.4 · p. 18
  • At 9B, online RL raises success from 45.6% to 50.8%; at 27B, it rises from 62.7% to 70.8%, indicating gains beyond supervised initialization at both scales.Both SFT and RL variants improve with model size, suggesting RUIC-SFT and DAO-GRPO convert increased capacity into stronger browser-agent capability.Source: §7.4 · p. 18
  • 70.8% average success rate leads open models, ahead of Qwen3.8-Max (70.3%); Wuying-Browser-Agent-27B scores 80.6% on WebVoyager, 66.7% on Online-Mind2Web, and 65.1% on BrowserBench.These results establish a new open-source state of the art across the three benchmarks.Source: §7.4 · p. 18, Table 2 · p. 19

7.5 Transfer to General Agentic Benchmarks

Wuying-Browser-Agent-27B transfers beyond browser use: it improves over Qwen3.5-27B across the reported general agentic benchmarks while remaining competitive with strong tool agents.

7.5 Transfer to General Agentic Benchmarks

  • Wuying-Browser-Agent-27B improves over Qwen3.5-27B on all reported Tau2-Bench, Claw-Eval, and BFCL-v4 benchmarks and remains competitive with strong general-purpose tool agents.The results indicate that browser-grounded training does not come at the expense of broader agentic capability.Source: §7.5 · p. 19
  • Wuying-Browser-Agent-27B is competitive overall with Qwen-UI-Agent-27B despite being trained primarily for real-web browser interaction.Source: §7.5 · p. 19
  • The reported transfer suggests gains in structured tool use, multi-turn task execution, and autonomous action under feedback.Source: §7.5 · p. 19

7.6 Ablation Study

The ablations use a 9B backbone on BrowserBench to isolate RUIC-SFT and DAO-GRPO components, and further assess estimator reliability, online training dynamics, performance slices, and browser-interaction behavior.

7.6.1 Ablation on RUIC-SFT

  • The full RUIC-SFT curriculum (#5) achieves 38.0% overall success, 18.5% recovery success, and 22.4 average steps per completed task.Adding Du raises overall success from 32.0% to 34.9% while recovery success changes from 8.5% to 9.0%; adding Dr raises recovery success from 8.5% to 16.4%. Fixed mixing raises steps from 24.8 to 29.3, while aggressive early reflection degrades overall performance versus fixed mixing and the curriculum.Source: §7.6.1 · p. 20, Table 4 · p. 21

7.6.2 Ablation on DAO-GRPO

  • 42.9% overall, 30.5% on hard tasks, and 32.8% on recovery-oriented tasks with DAO-GRPO’s full response-level objective under reconstructed contexts.PBRS raises overall success to 40.7%; divergence-aware credit improves hard-task success from 25.7% to 27.6% and recovery-oriented success from 26.2% to 29.7%. The components address sparse supervision, decisive branch steps, and decision-context alignment, respectively.Source: §7.6.2 · pp. 20–21, Table 5 · p. 21

7.6.3 Reliability of the Divergence Estimator

  • The divergence estimator localizes the human-labeled step exactly in 61% of groups, within one step in 78%, and within two steps in 89%.Branch partition agrees with human annotation in 94% of groups; perturbing the detected step by ±2 costs 0.9 points (42.9% → 42.0%), while ±4 retains 41.1%, above uniform credit at 40.7%. Random placement yields 40.3%, and hard-task success falls from 30.5% to 25.7%.Source: §7.6.3 · pp. 21–22, Figure 7 · p. 22

7.6.4 Reliability of the Progress Estimator Φ

  • On 150 held-out prefixes, Φ correlates with human progress ratings at Spearman ρ = 0.81, with mean absolute error 0.11 on the unit interval.It agrees with human ordering on 86% of 200 same-task prefix pairs and is monotone non-decreasing across 87% of adjacent steps; bounded potential-based shaping preserves the optimal policy despite estimator noise.Source: §7.6.4 · p. 22

7.6.5 Online RL Training Dynamics

  • During online training, valid-rollout validation reward reaches 0.76 versus raw validation reward 0.73 at iteration 150, while invalid-action ratios fall on both environments.Invalid actions decrease from 15.1% to 6.1% in training and 16.4% to 3.1% in validation. Zero-variance groups average 26.7% and persist; dynamic sampling masks them, while higher sandbox concurrency improves final performance by increasing data freshness and the pool available for sampling.Source: §7.6.5 · pp. 22–23, Figure 8 · p. 23

7.6.6 Performance by Task Difficulty and Trajectory Length

  • DAO-GRPO’s gain grows with difficulty and trajectory length, reaching +12.4 points on hard tasks and +13.4 points on tasks exceeding 50 steps.Hard-task success rises from 18.1% to 30.5%, and success on tasks exceeding 50 steps rises from 13.3% to 26.7%. RUIC-SFT improves all difficulty levels and trajectory-length bins, with a +22.2-point gain on 15–25-step tasks; longer tasks remain more challenging overall.Source: §7.6.6 · pp. 23–24, Figure 9 · p. 24

7.6.7 Case Study

  • Three rollout cases illustrate DAO-GRPO’s mechanisms: divergence-aware credit favors successful branches, PBRS orders mostly failing rollouts, and signed increments distinguish efficient from wasteful paths.In the first group, 12 of 16 top-down browsing rollouts succeed while 3 of 4 anchor-jumping rollouts stall at partial credit. In another, 16 of 17 rollouts fail, yet progress increments preserve useful ordering; in a third, equal-reward executions span 21–49 steps, with one receiving fourteen consecutive detour penalties.Source: §7.6.7 · pp. 24–25, Figure 10 · p. 25

7.6.8 Qualitative Analysis

  • A 15-step privacy-policy retrieval trajectory shows the agent detecting a misleading redirect, recovering, and switching from repeated scrolling to structured probing.The agent verifies browser state against its goal, notices the URL and content mismatch, re-enters the privacy path, and completes the requested inspection. The case illustrates closed-loop interaction under changing contexts rather than merely issuing actions and assuming progress.Source: §7.6.8 · pp. 25, 27, Figure 11 · p. 26

8 Conclusion

Wuying-Browser-Agent co-designs supervision, optimization, and evaluation for long-horizon browser deployment, addressing the mismatch between short-horizon training and real-world use. Experiments show the resulting series establishes a new open-source state of the art across multiple live-web benchmarks while remaining competitive with proprietary agents.

8 Conclusion

  • The Wuying-Browser-Agent series establishes a new open-source state of the art across multiple live-web benchmarks while remaining competitive with proprietary agents.The conclusion does not provide specific benchmark scores.Source: §8 · p. 27
  • The pipeline addresses the mismatch between success-dominated short-horizon training and real long-horizon browser deployment through co-designed supervision, online optimization, and evaluation.The paper identifies this mismatch as the central bottleneck limiting real-world web-agent robustness.Source: §8 · p. 27
  • RUIC-SFT combines reflection-rich recovery data and specialized UI interaction data through a progressive curriculum to build execution stability and self-correction priors.It provides a capability-structured initialization.Source: §8 · p. 27
  • DAO-GRPO refines long-horizon decisions in live browser interaction using potential-based reward shaping, divergence-aware step credit assignment, and response-level optimization under dynamically reconstructed contexts.Source: §8 · p. 27
  • BrowserBench supports fine-grained validation beyond aggregate success rates through structured success criteria and capability-aligned semantic tags.The conclusion describes it as a co-designed diagnostic instrument.Source: §8 · p. 27

Related papers

Loading 2608.17319v1…