TimeWarp: Evaluating Web Agents by Revisiting the Past

TL;DR

TimeTraj is proposed, a simple yet effective algorithm that uses plan distillation to collect trajectories across multiple versions to improve the robustness of web agents and unlock a new paradigm for collecting plans rather than trajectories, thereby improving the robustness of web agents.

Abstract

The improvement of web agents on current benchmarks raises the question: Do today's agents perform just as well when the web changes? We introduce TimeWarp, a benchmark that emulates the evolving web using containerized environments that vary in UI, design, and layout. TimeWarp consists of three web environments, each with six UI versions spanning different eras of the internet, paired with a set of complex, realistic tasks requiring different forms of web navigation. Our experiments reveal web agents' vulnerability to changes and the limitations of behavior cloning (BC) on single-version trajectories. To address this, we propose TimeTraj, a simple yet effective algorithm 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: $20.4\%\rightarrow37.7\%$ for Qwen-3 4B and $0\%\rightarrow27.0\%$ for Llama-3.1 8B models. We hope our work helps researchers study generalization across web designs and unlock a new paradigm for collecting plans rather than trajectories, thereby improving the robustness of web agents.

Figures (34)

• Figure 1: Dynamicity of the Web. Websites change visually and functionally over time, resulting in workflow changes.
• Figure 2: Overview of the TimeWarp benchmark: environments, versions, version to year mapping, UI examples, task distribution, and goal examples. The benchmark contains 231 goals $\times$ 6 versions $=$ 1386 tasks.
• Figure 3: Overview of the TimeWarp benchmark: (1) The goals and desired outcomes of the tasks are passed to the planner, $\Pi_\text{plan}$, which (2) produces draft execution plans. (3) The plans are refined by humans and (4) passed to the executor (teacher), $\Pi_\text{T}$, along with the goals and desired outcomes. (5) The executor generates trajectories across different versions of the TimeWarp environments: Wiki, News, and Shop. (6) Rollouts of the trajectories consist of observations and a 4-tuple of response tokens ($<$action$>$, $<$thinking$>$, $<$memory$>$, $<$plan$>$), at each time step. (7) Trajectories are evaluated by the judge $J_\phi$, and only the correct trajectories are filtered. (8) The filtered rollouts across versions form the training data for a web agent $\pi_{\theta}$, (9) which uses behavior cloning to produce a TimeWarp agent, $\pi_{\theta_{\text{TW}}}$.
• Figure 4: Success rate (%) of TW models (a) trained only on v$_6$ trajectories $\mathcal{D}_{\tau,6}$vs. continually on v$_6$ then v$_1$ trajectories $\mathcal{D}_{\tau,6\rightarrow1}$, (b) training with and without WebArena training data. In both cases, the agents are evaluated on the v$_6$ environments.
• Figure 5: Popup Ads in Shop v$_5$ affecting a web agent's workflow and task success.