Text-based web agents offer computational efficiency for autonomous web navigation, yet developing robust agents remains challenging due to the noisy and heterogeneous nature of real-world HTML. Standard Supervised Fine-Tuning (SFT) approaches fail in two critical dimensions: they lack discrimination capabilities to reject plausible but incorrect elements in densely populated pages, and exhibit limited generalization to unseen website layouts. To address these challenges, we introduce the Triton dataset (590k instances) and a progressive training curriculum. Triton is constructed via Structural-Semantic Hard Negative Mining, which explicitly mines topologically similar distractors, and a Dual-Agent Consensus pipeline that synthesizes diverse cross-domain tasks with strict verification. Building upon this foundation, our progressive curriculum produces three models: Triton-SFT-32B for basic imitation, Triton-ORPO-32B for robust discrimination via Odds Ratio Preference Optimization, and Triton-GRPO-32B for long-horizon consistency through Group Relative Policy Optimization. Empirical evaluation on Mind2Web demonstrates that Triton-GRPO-32B achieves state-of-the-art performance among open-source models with 58.7% Step Success Rate, surpassing GPT-4.5 (42.4%) and Claude-4.5 (41.4%) by over 16%, validating that specialized data curriculum outweighs raw parameter scale for web navigation.
Figures & tables
Figure 1: Standard imitation learning hallucinates on structurally similar elements (incorrectly selecting iPhone 17 Pro when instructed to buy iPhone 16 Pro), while Triton trained with progressive curriculum (ORPO for discrimination + GRPO for consistency) achieves target identification and multi-step task completion.
Figure 2: Overview of our proposed framework. (Left) We construct the Triton dataset by mining topological hard negatives and synthesizing cross-domain tasks via Dual-Agent Consensus. (Middle) This data fuels a progressive curriculum that evolves the model from imitation (SFT) to discrimination (ORPO) and consistency (GRPO), (Right) enabling robust execution in multi-step environments.
Component
Method
# Samples
Mind2Web (Base)
-
7k
+ Hard Neg.
DOM-Tree Mining
140k
+ Rejection
Perturbation
153k
In-Domain Total
300k
WebSight (Synth.)
Dual-Agent
290k
Triton Total
590k
Table 1: Composition of the Triton Dataset
Figure 3: Visualization of semantic diversity via t-SNE. While the original Mind2Web data (Blue) forms dense, domain-specific clusters, our synthetic WebSight data (Green) spans a broader semantic space. This complementary distribution demonstrates that Triton effectively fills the "generalization gaps" left by standard behavioral cloning datasets.
Split
# Tasks
# Domains
# Websites
Avg # Actions
Train
1,009
17
73
7.7
Cross-Task
177
17
64
7.6
Cross-Website
142
9
10
7.2
Cross-Domain
694
13
53
5.9
Table 2: Statistics of the Mind2Web dataset. The benchmark is divided into three test splits to evaluate generalization difficulty. Note that the Cross-Domain split contains the largest number of tasks but the shortest average trajectory length.
Model
Size
All
Cross-Domain
Cross-Task
Cross-Website
Score
SR
Score
SR
Ele
Op
M.Ele
M.Op
Score
SR
Ele
Op
M.Ele
M.Op
Score
SR
Ele
Op
M.Ele
M.Op
Closed-Source LLMs
Claude-4-5-noThink
58.8
41.4
58.1
41.8
48.3
67.0
49.9
67.2
61.5
43.7
52.0
68.3
55.2
70.4
56.9
38.7
46.8
65.1
48.7
67.0
Gemini-2.0-Flash
53.1
35.0
52.6
35.2
43.1
61.2
44.5
61.5
55.7
37.1
46.5
63.4
48.8
64.2
51.0
32.8
41.2
59.5
43.1
60.2
Gemini-2.5-Flash-preview-04-17
55.3
38.0
54.9
38.4
45.4
63.5
46.8
63.9
57.7
40.1
48.2
65.1
50.5
66.8
53.3
35.6
43.5
61.8
45.2
62.5
Gemini-2.5-pro-preview-05-06
59.1
41.5
58.3
41.6
48.5
67.2
50.1
67.5
61.7
43.9
52.2
68.5
55.4
70.6
57.2
39.1
47.1
65.4
49.0
67.2
Table 3: Main results on the Mind2Web benchmark. We report the Composite Score and Step Success Rate (SR) across three generalization splits. Triton-GRPO-32B significantly outperforms all open-source baselines (including the 480B model) and surpasses top-tier proprietary models like GPT-4.5 and Claude-4.5 in Step SR, demonstrating robust generalization despite its smaller parameter size.
Figure 4: Training dynamics analysis. Performance trajectories of the Full Curriculum versus variants with insufficient SFT foundations. Shaded areas denote alignment stages. The persistent gap confirms that ORPO and GRPO act as behavioral refiners rather than knowledge injectors, and thus cannot recover from a weak SFT baseline.
Model Variant
SFT Triton Data
RL Setting
Overall
Cross-Domain
Cross-Task
Cross-Website
Mind
Dis
Vis
ORPO
GRPO
Score
Step SR
Score
Step SR
Score
Step SR
Score
Step SR
Qwen2.5-Coder-32B-Instruct
✗
✗
✗
✗
✗
28.7
14.6
30.1
15.5
29.8
16.1
26.2
27.9
+ Mind
✓
✗
✗
✗
✗
55.1
36.3
55.1
36.3
57.5
38.8
52.8
33.8
+ Dis
✗
✓
✗
✗
✗
60.7
42.0
59.8
41.9
64.8
45.6
57.7
38.6
+ Vis
✗
✗
✓
✗
✗
39.6
21.5
39.3
21.2
40.9
22.5
38.6
20.8
+Dis + Vis
✗
✓
✓
✗
✗
60.5
47.6
58.2
45.2
65.5
53.1
57.8
45.0
Table 4: Component-wise ablation study on the Mind2Web development set. We incrementally validate the effectiveness of our data construction strategies (Mind vs. Dis vs. Vis) and training stages (SFT → ORPO → GRPO). Mind : Original training set; Dis : Discriminative Trajectory Mining data; Vis : Synthetic Visual-Structural Grounding data. Note that Triton-GRPO-32B combines all strategies to achieve optimal performance.
Figure 5: Hyperparameter sensitivity. Step Success Rate plotted against negative pool size ( K ) and rejection volume ( G ). The distinct peak at K=20,G=10 (Red line) indicates an optimal trade-off, demonstrating that performance degrades under both sparse discrimination signals ( K=1 ) and excessive noise ( G=50 ).
Figure 6: Scaling analysis of synthetic data. We measure performance across varying data sizes (0–29k). The observed monotonic growth, particularly the +3.3% boost in Cross-Domain (Blue), confirms that exposing the agent to diverse synthetic layouts serves as an effective structural regularizer for unseen websites.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Visualizing Hard Negatives in a flight search list. The model must discern the correct button based on the associated airline context, despite high structural similarity.
Figure 8: Visualization of Counterfactual Rejection. While baseline models often force an action on disabled elements due to instruction bias, Triton correctly outputs a rejection token.
Figure 9: Prompt templates used in the Dual-Agent Consensus pipeline. The Generator ( MG ) is conditioned on the target to create diverse intents, while the Verifier ( MV ) is blind to the target to ensure rigorous quality control.
Figure 10: The flattened input representation. We inject explicit IDs (e.g., id="42" ) into the HTML to enable precise referencing. The History block enables the model to maintain context across long horizons.
Hyperparameter
Value
General Configuration
Base Model
Qwen2.5-Coder-32B-Instruct
Precision
bfloat16
Optimizer
AdamW
Weight Decay
0.01
LR Scheduler
Cosine
Appendix
Table 5: Hyperparameter Settings. We list the specific configurations for SFT, ORPO, and GRPO stages. Note that the learning rate decays according to a cosine schedule across all stages.
Component
Value
Condition / Description
Format Reward ( Rfmt )
+0.1
Awarded if the output strictly follows the JSON/ChatML structure (valid Element ID and Operation fields).
Option Gate ( Iopt )
{0,1}
Acts as a multiplier. 1 if the predicted Element ID matches the ground truth, 0 otherwise.
Element Reward ( Ropt )
+1.0
Major reward for correctly identifying the target DOM element.
Arg F1 Score ( RF1 )
[0,1.0]
Token-level F1 score between predicted typed text (e.g., "iphone 13") and ground truth. Only calculated if Iopt=1 .
Perfect Bonus ( Rperf )
+1.0
Extra bonus if both Element and Operation arguments are perfectly correct ( F1=1.0 ).
Max Total Reward
3.1
0.1+1⋅(1.0+1.0+1.0)
Appendix
Table 6: Hierarchical Reward Breakdown. The reward function is designed as a cascade: the agent must first satisfy the format, then the element selection, and finally the operation arguments. If the element is wrong ( Iopt=0 ), subsequent rewards are zeroed out to prevent reward hacking.
As web agents close the gap with humans on benchmarks, one question arises: Do today's agents perform just as well on tomorrow's web? We introduce TimeWarp, a benchmark that emulates the evolving web. TimeWarp consists of three web environments, each with six UI versions spanning UI design, frontend code, and workflows from different eras of the internet. We pair TimeWarp with a set of complex, realistic tasks covering different forms of web navigation. Our experiments reveal that vision-based agents are vulnerable to changes, while text-based agents become brittle once fine-tuned on a single version. To address this, we propose TimeTraj, a new annotation method that uses plan distillation to collect trajectories across multiple versions. By training agents on teacher rollouts using our BC-variant, we achieve substantial performance gains: 20.4% to 37.7% for Qwen-3 4B and 0% to 27.0% for Llama-3.1 8B models. Our work helps study generalization across web designs and opens a new paradigm for collecting plans rather than trajectories to improve the robustness of web agents.
Frontier LLMs can navigate complex websites, but their cost and reliance on third-party APIs make local deployment impractical. We introduce Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents by analogy to human annotation roles, replacing the Task Designer, Annotator, and Supervisor with modular LLM components. Using Gemini 3 Pro as teacher, we generate 3,000 trajectories across six web environments and fine-tune a 9B-parameter student with pure supervised learning on the 2,322 that pass quality filtering. The resulting model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol, and nearly doubling the previous best open-weight result (Go-Browse, 21.7%). Capabilities transfer to unseen environments, with an 18.2 percentage point gain on WorkArena L1 (an enterprise platform never seen during training) and consistent improvements across three additional benchmarks. Ablations confirm that each pipeline component contributes meaningfully, with Judge filtering, evaluation hints, and reasoning traces each accounting for measurable gains. These results demonstrate that structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. Project page: https://agent-as-annotators.github.io
Large language models (LLMs) have enabled web agents that follow natural language goals through multi-step browser interactions. However, agents fine-tuned on specific trajectories and domain often struggle to generalize out of domain, and offline training can be compute-inefficient due to noisy, redundant trajectories and long accessibility-tree (AXTree) states. To address both issues, we propose Weasel, a trajectory selection method for offline training of web agents. Weasel selects a fixed-budget subset of trajectory steps by optimizing an objective that balances unary importance with pairwise diversity over states, websites, and interaction patterns, solving efficiently with a greedy algorithm. We further improve efficiency with target-centered AXTree pruning that keeps only content around the ground-truth action target, and we mitigate style mismatch for reasoning-native models by replacing expert traces with model-generated, style-consistent rationales. Across AgentTrek and NNetNav training datasets, evaluations in WebArena, WorkArena, and MiniWob, and experiments with Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel improves out-of-domain performance while reducing training cost, producing roughly 9.7-12.5× training speedups over standard fine-tuning. We make the code available at https://github.com/fatemehpesaran310/weasel.
Fatemeh Pesaran Zadeh, Seyeon Choi, Xing Han Lù +2
Seoul National University · McGill University · Mila – Quebec AI Institute +1