# Structured Distillation of Web Agent Capabilities Enables Generalization Source: [Structured Distillation of Web Agent Capabilities Enables Generalization]() Collection index: [llms.txt]() Data updated: 2026-04-13T06:00:00.940Z Authors: Xing Han L\\\`u, Siva Reddy Published: 2026-04-09T04:04:15.000Z ## Abstract 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 ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Enterprise interface interaction on WorkArena L2 (test)](): Enterprise interface interaction; WorkArena L2 (test); Success Rate: 9.7. - [Web navigation / Agent interaction on WebArena full 812-task](): Web navigation / Agent interaction; WebArena full 812-task; Success Rate: 42.1. - [Visual Web Navigation and Interaction on VisualWebArena](): Visual web navigation / Agent interaction; VisualWebArena; Success Rate: 0.337. - [Web Agent Navigation on WebArena (test)](): Web Agent Navigation; WebArena (test); Success Rate: 41.5. - [Web Agent Navigation on VisualWebArena (test)](): Web Agent Navigation; VisualWebArena (test); Success Rate: 33.9. - [Web Navigation on WorkArena L1](): Web navigation; WorkArena L1; Success Rate: 51.5. - [Web Agent Navigation on MiniWoB (full)](): Web Agent Navigation; MiniWoB (full); Success Rate: 69. - [Enterprise Interface Task Completion on WorkArena++ L2](): Enterprise interface task completion; WorkArena++ L2; Success Rate: 9.7. - [Web Agent Navigation on WorkArena L1 (full)](): Web Agent Navigation; WorkArena L1 (full); Success Rate: 51.5. - [Web Agent Navigation on WorkArena L2 147-task (test)](): Web Agent Navigation; WorkArena L2 147-task (test); Success Rate: 9.7. - [Web-based Interaction on MiniWoB](): Web-based interaction; MiniWoB; Success Rate: 69. - [Enterprise Interface Task Completion on WorkArena L1](): Enterprise interface task completion; WorkArena L1; Task Success Rate: 51.5. - [Web Navigation on VisualWebArena](): Web navigation; VisualWebArena; Published VWA Success Rate: 33.9. - [Web task automation on VisualWebArena full](): Web task automation; VisualWebArena full; SR: 33.9. - [Web Navigation on WebArena](): Web navigation; WebArena; Overall Success Rate: 41.5. - [Enterprise interface interaction on WorkArena L2 full benchmark](): Enterprise interface interaction; WorkArena L2 full benchmark; Success Rate: not recorded. ## Original sources and implementation [Original paper]() [PDF]() [Implementation]()