NexForge generates agent training data, beats Claude Opus 4.6 on benchmarks

A new paper from researchers introduces NexForge, a requirement-first framework for scaling executable agent training data. Traditional substrate-first methods tie task generation to predefined tools, repositories, or skill graphs, requiring manual expansion for each new domain. NexForge reverses this: it first performs research-based demand discovery to identify representative task forms and realistic scenarios, then compiles free-form capability requirements into executable training data. It automatically retrieves or constructs files, repositories, dependencies, and runtime configurations, followed by teacher rollout collection and trajectory distillation.

Experiments demonstrate the power of this approach. Using the same pipeline without any domain-specific infrastructure, NexForge generated 3,600 terminal tasks and 2,000 office tasks, improving Qwen3.5-35B-A3B Base from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval. Scaling to 43.2K terminal tasks pushed performance to 58.4%, surpassing Claude Opus 4.6. Furthermore, the Nex-N2 model family, trained with NexForge-synthesized data, lifts Qwen3.5-35B-A3B to 75.3% on Terminal-Bench 2.1 and 1585 Elo on GDPval, achieving state-of-the-art open-source performance and surpassing several proprietary systems.