NexForge Framework Scales LLM Agent Training by Auto-Generating Real-World Tasks

Training agents that can execute complex tasks remains expensive and hard to scale. Most existing methods bind task generation to predefined tools or repositories, meaning each new domain requires custom infrastructure and manual tweaking. Researchers have now introduced NexForge, a framework that sidesteps this bottleneck by synthesizing diverse, executable agent tasks directly from high-level capability requirements.

How NexForge Works#

Rather than relying on fixed toolsets, NexForge takes abstract capability requirements as input and then automatically generates both task directives and expert trajectories suitable for supervised fine-tuning. The process unfolds in stages:

• Demand analysis: The framework investigates real-world usage patterns to identify representative scenarios and task profiles.
• Distribution-aware compilation: It generates task directives that reflect actual demand patterns rather than substrate biases.
• Runtime assembly: For each task, NexForge automatically provisions the necessary files, dependencies, and runtime environments.
• Expert synthesis: It then produces expert rollouts and complete training trajectories ready for model training.

The key advantage is that this pipeline requires no domain-specific engineering—the same framework can target different capability areas without custom pipelines for each one.

Results and Performance Gains#

Testing on terminal and office automation tasks demonstrated substantial improvements. Starting with Qwen 3.5-35B-A3B as a baseline, NexForge's initial 3.6K terminal and 2K office tasks improved the model from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval. Scaling further to 43.2K terminal tasks pushed performance to 58.4%—matching Claude Opus 4.6 with Claude Code, a proprietary system.

Broader training on NexForge-generated data led to Nex-N2, a family of open-source agent models that achieved 75.3% on Terminal-Bench 2.1 and 1585 Elo on GDPval. These results represent state-of-the-art performance among publicly available systems and exceed several frontier proprietary alternatives.

Why This Matters#

The framework addresses a critical scaling challenge: as LLM agents move into production, generating training data becomes a significant bottleneck. By automating both task synthesis and scenario construction, NexForge decouples capability improvements from manual engineering overhead. This suggests a path toward scalable agent training that can adapt to new domains without specialized infrastructure—a step toward more flexible and capable autonomous systems.