Glonce

Mage-VL-Cowork, a 4B computer-use agent, reports 66.0% on OSWorld-Verified

agentsThe Glonce editorsarXiv🔥60Glonce rating60Impact45Novelty60Relevance100Surprise35Players10Credibility75

The paper introduces Mage-VL-Cowork, a 4B-parameter computer-use agent (CUA). Its starting point is the authors' view that effective CUAs should combine visual GUI interaction with code and terminal execution inside a single policy, rather than treating them as separate tools. They call the result a hybrid CUA.

The model is trained from a general multimodally aligned checkpoint through an end-to-end, progressive five-stage curriculum that spans midtraining, supervised fine-tuning (SFT) and reinforcement learning (RL).

To feed that curriculum at scale, the authors built Mage-CUA-Forge, a task-synthesis pipeline. Instead of constructing an agent for every task, it uses reusable, skill-conditioned generation workflows and application-specific preparers. It synthesized 293,806 desktop and web tasks across 6,915 skills, and the authors say it cut recurring per-task expenditure by 82.6%.

On the supervision side, the authors do not simply distill a default teacher. They evolve the teacher harness and instantiate specialized CUA harnesses, which yield high-quality trajectories together with stage-specific supervision. The title's phrase "evolved harnesses" refers to this.

The headline result: Mage-VL-Cowork-4B climbs to 66.0% on OSWorld-Verified across the training stages, with what the authors call more efficient reasoning cost. They describe this as state-of-the-art performance at its parameter size and as pushing forward the frontier of on-device intelligence. These are the authors' own assessments, stated in the abstract.

Key facts

  • Mage-VL-Cowork is a 4B hybrid computer-use agent that combines visual GUI interaction with code and terminal execution in one policy.
  • It is trained from a general multimodally aligned checkpoint through a five-stage curriculum across midtraining, SFT and RL.
  • The Mage-CUA-Forge pipeline synthesized 293,806 desktop and web tasks across 6,915 skills and, per the authors, cut recurring per-task expenditure by 82.6%.
  • Rather than distilling a default teacher directly, the authors evolve the teacher harness and build specialized CUA harnesses to produce trajectories and stage-specific supervision.
  • The authors report 66.0% on OSWorld-Verified and claim state-of-the-art performance at this parameter size.

Why it matters

The paper argues for a specific design: one small policy that clicks through a GUI and also writes code and runs terminal commands, instead of a GUI-only agent. It pairs that with a full training recipe, from the task-synthesis pipeline to the five training stages. A 4B model scoring 66.0% on OSWorld-Verified, if the result holds, would fit the authors' framing of on-device computer use.

Who it affects

Teams building computer-use agents, especially those who want small models they could run locally rather than large hosted ones. Researchers working on synthetic task generation and teacher-based supervision for agents will also find the Mage-CUA-Forge and evolved-harness ideas relevant.

How to use it

The abstract describes a recipe, not a product. The practical takeaways are the method's building blocks: start from a multimodally aligned checkpoint, move through midtraining, SFT and RL, generate tasks with reusable skill-conditioned workflows instead of per-task agents, and evolve the teacher harness rather than distilling a default teacher. The abstract does not say whether the model, data or code are released.

How solid is it

This is an arXiv abstract, and every figure in it is the authors' own report. The 66.0% on OSWorld-Verified is a concrete number on a named benchmark. The state-of-the-art claim at this size and the claim of more efficient reasoning cost are self-assessments. No baseline or competitor scores on OSWorld-Verified are given, so the margin over other models of the same size is not stated. The abstract names no authors or institutions.

Risks and caveats

The 82.6% cost reduction is not tied to a stated baseline method, dollar amount or unit; only "recurring per-task expenditure" is named. No quantification is given for "more efficient reasoning cost". The abstract does not say which teacher model is used or what the five stages individually are, so the recipe cannot be judged in detail from this text alone.

“Effective computer-use agents (CUAs) should seamlessly combine visual GUI interaction with code and terminal execution within a single policy.”

— Mage-VL-Cowork paper abstract