Fara-1.5: Scalable Learning Environments for Computer Use Agents
FaraGen1.5 enables scalable data generation for computer use agents using modular components, achieving 72.3% on Online-Mind2Web with Fara1.5-27B.
Key Findings
Methodology
FaraGen1.5 consists of environments, solvers, and verifiers, using real and synthetic environments for data generation. The solver employs GPT-5.4, and verifiers ensure task correctness, efficiency, and critical-point adherence.
Key Results
- Fara1.5-9B achieves 63.4% on Online-Mind2Web and 86.6% on WebVoyager, significantly outperforming Fara-7B and GUI-Owl-1.5-8B.
- Fara1.5-27B reaches 72.3% on Online-Mind2Web, competitive with larger proprietary systems.
- Data generated by FaraGen1.5 sets new standards for browser-use benchmarks with the Fara1.5 series.
Significance
This research addresses the cost and speed issues of collecting data through human demonstrations, providing a scalable solution that advances computer use agent research and application.
Technical Contribution
FaraGen1.5 offers a modular data generation framework, combining powerful solvers and verifiers to enhance data quality and model performance.
Novelty
First to combine synthetic and real environments, providing comprehensive task verification to ensure data quality and model training reliability.
Limitations
- Complexity of synthetic environments may limit realism for certain tasks.
- Verifiers may misjudge in certain scenarios.
- Models may underperform in unpredictable environments.
Future Work
Future work will focus on expanding the complexity and diversity of synthetic environments and improving verifier accuracy.
AI Executive Summary
FaraGen1.5 achieves scalable data generation for computer use agents through modular components. The system combines real and synthetic environments, using a GPT-5.4 solver and multiple verifiers to ensure task correctness and efficiency. The Fara1.5 series sets new standards in browser-use benchmarks, significantly improving task success rates. This research not only addresses the cost and speed issues of data collection but also advances computer use agent research and application. Future work will focus on expanding the complexity and diversity of synthetic environments and improving verifier accuracy.
Deep Analysis
Background
Computer use agents play a crucial role in modern technology, especially in browser task automation. Previous research relied heavily on human demonstration data, which is costly and slow.
Core Problem
Collecting high-quality computer use data is a bottleneck for training agents. Existing methods cannot effectively generate large-scale, diverse data.
Innovation
FaraGen1.5 achieves data generation through modular environments, solvers, and verifiers. Synthetic environments allow safe attempts of irreversible tasks, solvers use GPT-5.4, and verifiers ensure data quality.
Methodology
- �� Use real and synthetic environments for data generation
- �� Solver employs GPT-5.4 model
- �� Verifiers ensure task correctness, efficiency, and critical-point adherence
Experiments
Experiments use Online-Mind2Web and WebVoyager datasets, comparing Fara1.5 series with other models to validate data generation effectiveness.
Results
Fara1.5-9B achieves 63.4% on Online-Mind2Web and 86.6% on WebVoyager. Fara1.5-27B reaches 72.3% on Online-Mind2Web.
Applications
The system can automate browser tasks such as form filling and online shopping, enhancing user experience and efficiency.
Limitations & Outlook
Complexity of synthetic environments may limit realism for certain tasks, and verifiers may misjudge in certain scenarios.
Plain Language Accessible to non-experts
Imagine a virtual office where FaraGen1.5 acts like a super secretary, handling various tasks in real and simulated environments. It uses powerful solvers to tackle problems and verifiers to ensure accuracy at every step. Just like a secretary checks every detail when handling complex tasks, FaraGen1.5 ensures each data point is reliable.
ELI14 Explained like you're 14
Imagine playing a super game where FaraGen1.5 is your game assistant, helping you complete various tasks. It's like a smart robot working in real and simulated worlds, ensuring you win every time. It uses powerful tools to solve problems, just like you use super weapons in games.
Glossary
FaraGen1.5
A modular data generation framework for computer use agents.
Used for generating high-quality training data.
GPT-5.4
A powerful language model used in the solver.
Employed in the solver to generate task solutions.
Verifier
Component ensuring data quality and task correctness.
Used to evaluate whether generated data meets standards.
Synthetic Environment
Virtual scenarios simulating real environments for safe task attempts.
Used for generating data for irreversible tasks.
Fara1.5
A series of computer use agent models.
Sets new standards in browser-use benchmarks.
Open Questions Unanswered questions from this research
- 1 How to further enhance the complexity and realism of synthetic environments?
- 2 How to improve verifier accuracy to reduce misjudgments?
Applications
Immediate Applications
Browser Task Automation
Use Fara1.5 models to automate browser tasks like form filling and online shopping.
Long-term Vision
Smart Office Assistant
Integrate Fara1.5 into smart office systems to enhance work efficiency and user experience.
Abstract
Collecting computer use data from human demonstrations is expensive and slow, motivating the need for scalable generation strategies. This requires two key ingredients: environments in which agents can act and verifiers that can judge whether their demonstrations succeeded. We introduce FaraGen1.5, a scalable data pipeline for computer use agents composed of three modular components: environments, solvers, and verifiers. FaraGen1.5 uses both live websites and synthetic environments that faithfully simulate domains gated by authentication or that require irreversible actions. It employs a solver harness that can be powered by multiple models, including strong frontier models such as GPT-5.4, and also incorporates a user simulator to enable multi-turn rollouts. Finally, FaraGen1.5 scores the resulting trajectories with three complementary verifiers covering task correctness, efficiency, and critical-point adherence. Using data produced by this pipeline, we train Fara1.5, a family of native computer use agents (CUAs) at three scales built on Qwen3.5 (4B, 9B, and 27B). To train these models, we employ a supervised finetuning (SFT) recipe that carefully balances data from FaraGen1.5 for broad coverage, specific high-value tasks, and target model deficiencies in an iterative approach. Each model sets a new state of the art (SoTA) for its size class on browser-use benchmarks: Fara1.5-9B reaches 63.4% on Online-Mind2Web and 86.6% on WebVoyager, while Fara1.5-27B achieves 72.3% on Online-Mind2Web, which is competitive with much larger proprietary systems. We also release weights for the Fara1.5 models under MIT license, making SoTA computer use accessible for all beyond closed API-only systems.