Back to Basics: A Skill-Centric Approach to Generalist Web Agents

Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui, Reaz Hassan Joarder, Kenneth MarinoarXiv 2026 · arXiv

TL;DR

This paper argues that generalist web agents should be built from the bottom up by mastering reusable "interaction primitives" (like clicking a slider or filling a form) rather than just practicing end-to-end tasks. The authors introduce MiniWeb, a diagnostic benchmark that reveals "drag & gesture" as a universal weakness in current models, and MiniWebAgent, a planner-executor system that achieves significant performance gains on unseen websites by composing these fundamental skills.

What Problem Does This Paper Address?

Building autonomous agents that can navigate the web is a major goal in AI, but current models often fail in ways that are hard to diagnose. Most existing benchmarks only provide a "pass/fail" score for entire tasks, leaving researchers in the dark about whether an agent failed because it couldn't find a button, misunderstood the instructions, or simply couldn't operate a slider.

Furthermore, training agents on massive amounts of end-to-end task data often leaves them "brittle." They might learn to solve specific workflows on popular sites but fail to generalize to new environments. The authors identify a specific "skill gap": even advanced open-weight models are surprisingly bad at basic physical interactions like dragging, which are rarely represented in standard training sets.

Methods

The authors propose a "Back to Basics" approach consisting of three main components:

  1. MiniWeb Benchmark: A miniature web environment of 65 interconnected sites. Unlike other benchmarks, every task in MiniWeb is decomposed into 37 interaction primitives (e.g., filter_by_slider, sort_by_form). This allows for fine-grained "skill profiling" to see exactly where an agent's capabilities break down.
  2. MiniWebGen: An automated pipeline that uses a "teacher" model with privileged access to website internals to generate thousands of verified, skill-centric training trajectories. This solves the data scarcity problem for rare but essential skills.
  3. MiniWebAgent: A modular agent architecture. It uses an untrained planner (a base LLM) to break a high-level request into a sequence of primitives, and a specialized executor (a fine-tuned compact model) to carry out each primitive.
The MiniWebAgent architecture uses a planner to decompose tasks into skills, which are then executed by a trained model.

Datasets

  • MiniWeb (New): 65 sites across 15 domains with 369 human-annotated tasks. It includes both public data (e.g., arXiv, Wikipedia) and synthetic data.
  • MiniWebGen Data: 3.1K verified trajectories generated for training the executor.
  • External Benchmarks (Evaluation): To test generalization, the authors evaluated their agent on WebArena-Lite, WebVoyager, and Online-Mind2Web without any training on those specific datasets.

Key Insights and Ideas

  • The "Drag" Bottleneck: Across ten different open-weight agents, "drag & gesture" was consistently the weakest skill family. Most agents failed not because they couldn't find the slider, but because they couldn't execute the precise physical movement required to set it.
Diagnostic profiling shows that dragging is a major failure point for all tested agents.
  • Failures Compound: The authors found that an agent's success on a multi-step task is roughly the product of its success on each individual step ($p^n$). This means even small improvements in basic skill reliability lead to massive gains in complex task success.
  • Composition is Key: Simply training a model on individual skills isn't enough; a single agent trained this way often "stops early" because it hasn't learned how to chain skills together. The explicit separation of planning and execution in MiniWebAgent is crucial for turning skill-level mastery into task-level success.
  • Efficiency of Skill-Centric Data: Training on 3.1K skill-focused trajectories outperformed training on a similar amount of task-centric data, suggesting that "practicing the right things" is more important than just practicing more.

Key Results

MiniWebAgent, built on a modest Qwen3.5-4B base, showed impressive improvements:

  • Task Success: Improved success by 19.4 percentage points (pp) on held-out MiniWeb sites.
  • Zero-Shot Transfer: Without seeing any training data from external benchmarks, it gained 9.3pp on WebArena-Lite, 20.0pp on WebVoyager, and 16.1pp on Online-Mind2Web.
  • Skill Mastery: Fine-tuning increased single-primitive success from 55.3% to 80.3% on held-out sites.
Success rates drop sharply as tasks require more primitives, highlighting the importance of skill reliability.

Limitations and Future Work

  • Human Task Scale: While MiniWeb is highly diagnostic, it currently has fewer human-authored tasks than massive benchmarks like Mind2Web.
  • Drag & Gesture Still Hard: Although the paper identifies the drag/gesture weakness and improves it, these skills remain the most difficult for agents to master completely. The authors suggest specialized methods for these interactions as a future direction.
  • Scaling: The study focused on compact open-weight models (4B-8B parameters). It remains to be seen how this skill-centric approach scales to much larger frontier models.

Summary updated

Abstract

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.
BibTeX
@misc{phamdinh2026back,
  author = {Minh Pham-Dinh and Md Farhan Ishmam and Hernan Yupanqui and Reaz Hassan Joarder and Kenneth Marino},
  title = {{Back to Basics: A Skill-Centric Approach to Generalist Web Agents}},
  year = {2026},
  eprint = {2610.35823},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35823},
}
Figure