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.
Figures & tables
Figure 1: Performance comparison on online browser-use benchmarks and general agentic benchmarks.
Figure 2: Three structural challenges for browser agents in the long-horizon regime, and our co-designed solution. (a) Training data is dominated by successful trajectories, but real long-horizon execution routinely involves mistakes and detours, leaving recovery and complex-UI behavior weakly supervised. (b) Long browser trajectories often share long prefixes and differ at only a few decisive branch steps, so uniform trajectory-level optimization dilutes the learning signal. (c) Existing benchmarks underrepresent bilingual long-horizon real-web tasks, leaving the deployment regime we target insufficiently evaluated.
Figure 3: Overview of the Wuying-Browser-Agent pipeline. A shared browser harness layer supports two-stage policy training, with RUIC-SFT learning from around 3K curated trajectories and DAO-GRPO refining the policy through over 5K live-web rollouts. BrowserBench provides bilingual long-horizon evaluation, while a self-reinforcing data flywheel feeds verified interaction outcomes back into subsequent training.
Table 1: Browser Agent Action Space. We define a structured browser action space with 24 atomic operations used in both SFT and online RL training. Actions are organized into five functional categories, including navigation, interaction, extraction, file operations, and flow control.
Figure 4: Piecewise linear annealing schedule for dataset mixing in RUIC-SFT. Training begins with a general-data-dominant distribution to stabilize basic browser operations, gradually increases the proportion of specialized UI component data to strengthen complex interaction skills, and introduces reflection data only in the late stage to cultivate self-correction without destabilizing the execution prior.
Figure 5: System overview of the online reinforcement learning stage. A task is sampled from the task pool D and assigned to the browser agent controller, which interacts with multiple AgentBay sandboxes in parallel. Each sandbox produces one rollout trajectory, and the resulting trajectory group is consumed by the DAO-GRPO optimization module. The same GPU backend supports both policy serving during rollout and parameter updates during online optimization.
Figure 6: Overview of BrowserBench. BrowserBench contains 350 bilingual real-web tasks spanning 254 websites, with an average completion length of 37.9 steps. The benchmark covers eight task categories across Chinese and English websites, and each case is normalized into a goal-only instruction paired with a structured success criterion, enabling reliable Pass@1 evaluation of long-horizon browser agents.
Model Name
WebVoyager
Online-Mind2Web
BrowserBench
Average
Closed-Source Models
GPT-4o
48.7
27.9
22.0
32.9
GPT-5
80.1
67.1
61.7
69.6
GPT-5.5
85.1
74.7
67.4
75.7
Ernie-5.0 (>1T)
51.9
31.3
24.9
36.0
Qwen3.7-Plus
73.6
60.1
59.4
64.4
Table 2: Task success rates (%) on three open-web benchmarks: WebVoyager, Online-Mind2Web, and BrowserBench. All models are evaluated with a maximum budget of 100 interaction steps. The Average column is the arithmetic mean of the three benchmarks. The best result in each column is shown in bold, and the second-best is underlined.
Model
Tau2-Bench
Claw-Eval (Avg 3)
Claw-Eval (Pass 3 )
BFCL-v4
Average
Browser Agents
OpenCUA 72B
14.4
26.4
0.5
28.3
17.4
GUI-Owl-1.5-32B-Instruct
6.1
29.6
5.5
32.7
18.5
UI-Venus-1.5-30B-A3B
22.7
30.6
5.5
19.8
19.7
EvoCUA-32B-20260105
48.9
46.3
6.5
48.8
37.6
General Models
Table 3: Agentic performance comparison on selected benchmarks. We report performance on Tau2-Bench, Claw-Eval, and BFCL-v4 to evaluate whether browser-grounded training preserves broader tool-use and autonomous-agent capability beyond browser-specific tasks.
#
Variant
Data Sources
Schedule
SR (%)
Recov. SR (%)
Steps
Dg
Du
Dr
1
Base SFT
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0.7852,0,0}\times}
{\color[rgb]{0.7852,0,0}\times}
None
32.0
8.5
24.8
2
+UI
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0.7852,0,0}\times}
Fixed
34.9
9.0
23.1
3
+Reflection
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0.7852,0,0}\times}
{\color[rgb]{0,0.5898,0}\checkmark}
Fixed
35.5
16.4
27.6
4
+UI+Reflection
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
Fixed
36.6
17.1
29.3
5
RUIC-SFT (full)
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
Curriculum
38.0
18.5
22.4
Table 4: Ablation on RUIC-SFT. All variants use the same 9B backbone and are evaluated on BrowserBench. “Fixed” applies a fixed multi-source mixture uniformly throughout training, while “Curriculum” uses the phased schedule. SR: overall task success rate; Recov. SR: erroneous-step recovery rate; Steps: average number of actions per completed task.
Variant
PBRS
Div. Credit
Resp.-Level
SR (%)
Hard SR (%)
Recov. SR (%)
Vanilla Online GRPO
{\color[rgb]{0.7852,0,0}\times}
{\color[rgb]{0.7852,0,0}\times}
{\color[rgb]{0.7852,0,0}\times}
39.4
23.8
24.8
+PBRS
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0.7852,0,0}\times}
{\color[rgb]{0.7852,0,0}\times}
40.7
25.7
26.2
+PBRS+Div. Credit
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0.7852,0,0}\times}
41.9
27.6
29.7
DAO-GRPO (full)
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
{\color[rgb]{0,0.5898,0}\checkmark}
42.9
30.5
32.8
Table 5: Ablation on DAO-GRPO components. All variants are initialized from the same RUIC-SFT checkpoint and trained for the same number of online iterations. SR: overall task success rate; Hard SR: success rate on hard BrowserBench tasks; Recov. SR: success rate on recovery-oriented tasks.
Figure 7: Reliability of the LLM-based divergence estimator. (a) Agreement between estimator outputs and human annotation on 100 labeled rollout groups: exact / ± 1-step / ± 2-step localization of t∗ and favorable–unfavorable branch partition. (b) BrowserBench success rate of DAO-GRPO (9B) when the detected t∗ is randomly perturbed by up to ± 2 or ± 4 steps or replaced by a random step; the dashed line marks the uniform-credit variant (+PBRS, 40.7%) from Table 5 .
Figure 8: Online RL training dynamics of Wuying-Browser-Agent-9B. (a) Shaped reward on training batches (per-step scatter with moving average) and on a held-out validation split, the latter also reported over valid rollouts only. (b) Invalid-action ratio on the training and validation environments throughout online RL. (c) Fraction of rollout groups with zero reward variance in each iteration; these groups contribute no relative-advantage signal and are masked by dynamic sampling. The dashed line marks the training average of 26.7%. (d) Final success rate under different training budgets (up to 150 iterations) and parallel sandbox concurrency levels. Overall, online optimization improves both reward and action validity, degenerate groups persist throughout training and are handled by dynamic sampling, and higher sandbox concurrency yields better final performance under the same number of update steps.
Figure 9: Success rate by task difficulty (a) and (b) trajectory length in steps on 9B model backbone. Trajectory length is binned into four ranges: 15–25, 26–35, 36–50, and > 50 steps.
Figure 10: Real rollout groups from online RL training, illustrating the three design choices of DAO-GRPO. For analysis we sample 16–20 rollouts per task (beyond the training group size K=6 ) and display six representative ones per group; row ends show final rewards, and the bottom strips show the mean per-step PBRS shaped reward of the displayed rollouts. (a) A Galaxy Z Fold7 browsing group: all rollouts share an identical three-step prefix (shaded corridor) and diverge at t∗=4 into top-down browsing ( B+ : scroll ) vs. anchor jumping ( B− : scroll_to_text ); 12 of 16 browsing rollouts succeed while 3 of 4 jumping rollouts stall at partial credit, and progress credit concentrates on the post-divergence steps of the favorable branch. (b) A Galaxy S23 Ultra group in which 16 of 17 rollouts fail with returns ≤0.06 : under terminal-only rewards the group would be degenerate (advantage ≈0 ), whereas PBRS progress increments keep the rollouts correctly ordered ( 1.00>0.06>0.03>0.00 ). (c) An accessory-filtering group in which 12 of 16 rollouts all reach reward 1.0 with lengths from 21 to 49 steps; signed PBRS increments penalize individual detour steps (crosses; deep dots: credited steps; light dots: no credit), e.g. 14 consecutive penalized steps in R5, separating efficient from wasteful executions that identical terminal rewards cannot distinguish.
Figure 11: Self-reflective error recovery in long-horizon web navigation. A 15-step trajectory of Wuying-Browser-Agent-9B on a Chery privacy-policy retrieval task. The agent detects an erroneous redirect from the observed browser state, performs corrective recovery, and escalates from scrolling to structured probing before reaching the personal information collection list. For ease of visualization, we display screenshots for all steps to highlight interface transitions; in actual deployment, the agent primarily consumes structured browser state (e.g., DOM-derived representations), while screenshots are used only as optional inputs when visual grounding is needed.
Browser agents do not just need more practice. They need to practice the right things. We introduce MiniWeb, a realistic miniature of the web comprising 65 interconnected websites across 15 domains, where agents navigate, search, and transact as people do. Its 369 human-annotated tasks cover everyday workflows and label 37 reusable interaction primitives grouped into 7 skill families, allowing us to diagnose which skills fail rather than only whether a task succeeds. Benchmarking ten open-weight agents, we find that drag & gesture is consistently the weakest family, with no agent solving more than 24% of its instances. To address this gap, we introduce MiniWebGen, a framework that autonomously generates verified, skill-centric training data from MiniWeb. Fine-tuning Qwen3.5-4B on just 3.1K verified trajectories improves single-primitive success by 24.9pp across 13 unseen websites, with gains across every skill family. However, these trajectories teach short-horizon interactions, which do not directly translate to full-task success. We build on this result with MiniWebAgent, which pairs the trained executor with a planner that decomposes full tasks into the skill-level steps it learned to perform. This approach improves task success by 19.4pp points on unseen MiniWeb sites and transfers to external benchmarks, with gains of 9.3pp on WebArena-Lite, 16.1pp on Online-Mind2Web, and 20.0pp on WebVoyager, without training on their trajectories. We provide a simple recipe for compact browser agents: practice reusable skills directly, then compose them to solve unfamiliar tasks.
Existing web agent benchmarks have largely converged on short, single-site tasks that frontier models are approaching saturation on. However, real world web use consists of long-horizon, multi-site workflows. Common web navigation tasks, such as comparing products across different domains, planning trips across multiple services, or summarizing information from multiple search queries, require sustained context and cross-site reasoning over potentially hours of browsing. To capture and evaluate such behaviors, we introduce Odysseys: a benchmark of 200 long-horizon web tasks derived from real world browsing sessions evaluated on the live Internet. We find that binary pass/fail evaluation is inadequate for long-horizon settings and introduce a rubric-based evaluation, annotating each Odysseys task with an average of 6.1 graded rubrics. We demonstrate that this yields higher agreement with humans and provides a more fine-grained signal than commonly used trajectory-level LLM-as-a-judge evaluation metrics. We tested several leading frontier models and find that the strongest models achieve a success rate of 44.5%, which leaves substantial room for future improvements. Beyond task success, we argue that efficiency is a first-class concern for long-horizon agents. We introduce a Trajectory Efficiency metric (rubric score per step) and find that even frontier agents achieve only 1.15%, marking an evident need for agents that can succeed efficiently and not simply eventually. Odysseys isolates the critical evaluation of long-horizon proficiency in open-web environments, providing a realistic benchmark to measure progress towards computer-use agents that can potentially productively operate for hours. We release our tasks, evaluation scripts, and other results at https://odysseys-website.pages.dev
Lawrence Keunho Jang, Jing Yu Koh, Daniel Fried +1
Large language models are increasingly deployed as autonomous agents that interact with the web through browsers. While recent progress has been driven by benchmarks that evaluate end-to-end task success, these evaluations largely overlook two fundamental sources of difficulty in real web browsing: complex actions over rich user interfaces and visual perception of dynamically rendered content, especially in workflows that span multiple websites. We introduce CAP, a scalable benchmark for evaluating browser agents on cross-site, human-like web tasks that require non-trivial UI interactions and visual understanding. Specifically, we adopt a decomposition-and-recomposition pipeline that first abstracts each website into a structured site card capturing user-facing functions, complex execution operations, and perceptual requirements, and then recomposes these components into realistic cross-site workflows. Each task is therefore grounded in multiple specific operations on each website, enabling fine-grained diagnosis. Built on this framework, we construct 420 tasks across 108 real-world websites and 24 domains under careful quality control. Experiments on state-of-the-art browser agents using our verifiable agent-as-a-judge evaluation framework show low success rates and reveal that perception-heavy interactions remain a major bottleneck, exposing substantial gaps between current agents and real-world web browsing demands.
Zejun Xu, Taiyi Chen, Jin Li +13
Macau University of Science and Technology · FellouAI · Tsinghua University +3