WebChallenger: A Reliable and Efficient Generalist Web Agent

Abstract: Autonomous web navigation remains challenging for LLM agents, and the strongest generalist systems rely on proprietary reasoning models whose inference cost is prohibitive for the repetitive tasks where such agents would be most useful. We argue this gap stems not from insufficient model capability but from agent architectures that fail to replicate three human cognitive advantages: selective attention to relevant page regions, persistent memory of website structure, and procedural fluency with common interaction patterns. We introduce WebChallenger, a web agent framework that addresses each gap through architecture design rather than model scale, built around PageMem: a structured page representation deterministically constructed from the DOM that exposes each page as a hierarchy of semantic sections with short summaries. On this shared substrate we build three mechanisms that mirror the three cognitive advantages: a divide-and-conquer observation pipeline that lets the agent skim section summaries and extract details only from task-relevant regions; a lightweight exploration and memory system that traverses each website once to build a reusable map of pages and element behaviors; and compound action workflows that collapse common multi-step interactions into single agent actions, handling partial state changes automatically. Because all three operate over PageMem, the framework generalizes across websites without site-specific adapters. Using off-the-shelf open-weight models without fine-tuning, our system achieves 56.3% on WebArena, 48.7% on VisualWebArena, 51.0% on Online-Mind2Web, and 70.9% on WorkArena, approaching frontier proprietary systems at a fraction of the cost. Our code is released at https://github.com/jayoohwang1/webchallenger

• The paper's main contribution is the introduction of architectural scaffolding, including PageMem, that enables efficient long-horizon web navigation using open-source LLMs.
• It employs a divide-and-conquer observation pipeline, persistent offline memory, and compound actions to drastically reduce context token bloat.
• Empirical results reveal state-of-the-art performance across multiple benchmarks, achieving near-proprietary efficiency at minimal inference cost.

Architectural Advances of WebChallenger: Reliable and Efficient Generalist Web Agent

Introduction and Motivation

"WebChallenger: A Reliable and Efficient Generalist Web Agent" (2606.10423) addresses persistent barriers in web navigation for LLM-powered agents, notably the inability of non-proprietary models to match specialized agents in long-horizon, realistic web tasks. The authors argue that previous gaps stem less from model limitations and more from failures in agent architecture: specifically, the lack of selective attention, persistent memory, and procedural fluency found in human users. The paper introduces WebChallenger, a system that achieves state-of-the-art performance among open-source agents without fine-tuning or reliance on proprietary models, relying instead on architectural scaffolding that generalizes across diverse sites.

Figure 1: WebChallenger achieves superior performance across four benchmarks, leveraging efficient architectural scaffolding with minimal compute and without model finetuning.

Agent Architecture: PageMem and Human-Inspired Mechanisms

Central to WebChallenger is the PageMem abstraction—a structured representation of each webpage, constructed deterministically from the DOM, partitioned into hierarchical semantic sections with short summaries. This enables token-efficient, region-focused processing that mimics human attentional fragmentation [10.1007/978-3-031-28238-6_5], contrasting the context dilution from flat token sequence ingestion.

Three mechanisms operate over PageMem:

• Divide-and-conquer observation pipeline: The agent initially skims section summaries, extracts detailed information only from task-relevant regions, and synthesizes concise, information-dense observations. This decomposition reduces context token bloat, enhances reasoning density, and facilitates robust scaling to small LLMs.
• Persistent memory via offline exploration: Before execution, deterministic site traversal builds WebsiteMem—persistent, structured maps of pages and element behaviors, enabling one-shot amortization of environmental knowledge, navigation shortcuts (bookmarks), and robust dropdown handling.
• Compound action workflows: The agent collapses common multi-step interactions (forms, dropdowns, search, etc.) into single agent actions with internal workflow execution, automatically surfacing partial state changes, minimizing redundant LLM calls, and aligning action granularity with semantic transitions.

  Figure 2: Architectural overview: DOM-based section decomposition, indexed summaries, and specialized workflows interplay to deliver scalable and generalizable web agent behavior.

Empirical Evaluation and Component Analysis

WebChallenger achieves 56.3% success on WebArena, 48.7% on VisualWebArena, 51.0% on Online-Mind2Web, and 70.9% on WorkArena—each a new state-of-the-art among open-weight agents. Notably, these results approach proprietary systems (e.g., OpenAI CUA, IBM CUGA, GPT-5) at a fraction of inference cost. All runs use off-the-shelf GLM-4-32B and Qwen VL backbones, with zero finetuning.

Ablation studies reveal:

• Removing the observation pipeline drops performance by 17.6 points, underscoring the impact of selective attention and region-based compression.
• Removing compound actions reduces accuracy by 9.7 points, especially in tasks with recurring interface complexities.
• Memory removal yields a smaller but non-trivial 7.6 point reduction.

Backbone sensitivity shows the architecture's portability: generic LLM harnesses score $19.4\%$ (GLM-4-32B) versus $58.8\%$ in WebChallenger, demonstrating a ~40 point gain from architectural scaffolding alone. Token and step efficiency analyses demonstrate that WebChallenger trades prompt complexity for step count—smaller, numerous prompts increase reasoning density without exceeding context limits.

Methodological Landscape and Comparative Positioning

WebChallenger advances several recent lines:

• Memory systems: Unlike workflow-memory and skill-learning approaches which accumulate task experience [wang2024agentworkflowmemory, zheng2025skillweaverwebagentsselfimprove], WebChallenger's exploration is deterministic and site-agnostic, requiring no demonstrations or external documentation.
• Observation refinement: Many agents prune HTML or vision contexts [gur2024realworldwebagentplanninglong, kerboua2025focusagentsimpleeffectiveways] or use pixel-space cropping, while WebChallenger preserves DOM semantics with lossless chunking and hierarchical indexing.
• Compound actions: Prior work extends action spaces with site-specific code or tool interfaces [song2025browsingapibasedwebagents, zhong2026actionenginereactiveprogrammaticgui], while WebChallenger's workflows are indexed to page structure, enabling generalization without per-site adaptation.
• Prompt optimization: The divide-and-conquer pipeline exploits context decomposition strategies and focused sub-prompts [zhang2026recursivelanguagemodels] enabling efficient scaling to small LLMs.

Implications, Limitations, and Future Directions

Practically, WebChallenger’s results imply that open-weight LLMs, with proper architectural scaffolding, are sufficient for many knowledge-work automation and evaluation tasks. This democratizes agent research, enabling deployment on private devices and reducing reliance on costly APIs. The framework is conducive to integration with advanced memory systems, skill abstraction, and synthetic-data generation. The authors note extension potential towards richer memories, online workflow learning, and adversarial robustness testing [tur2025safearenaevaluatingsafetyautonomous, zheng2025webguardbuildinggeneralizableguardrail].

Limitations include hand-coded structural priors (section decomposition, element identification, compound workflows) which may underperform on non-conforming websites, increased wall-clock time due to multiple LLM calls, and evaluation limited to benign task scenarios.

Conclusion

WebChallenger demonstrates that architectural innovation, rather than model scale or extensive fine-tuning, is the dominant constraint for open-weight web agents. By leveraging divide-and-conquer observation, persistent memory from deterministic exploration, and site-agnostic compound workflows on a shared DOM-derived substrate, it achieves competitive performance with proprietary systems and sets new standards for efficient, reliable web automation using small LLMs. This points toward a research frontier in compositional agent design where memory, attention, and procedural abstraction converge to match or surpass frontier models—even when running locally and untrained.