NexForge synthesizes agent training data from requirements and lifts Qwen3.5 by 30 points

nex-agi · hf · 2026-07-22

NexForge proposes a requirement-driven pipeline for synthesizing executable agent training data without hand-built substrates.

It starts from high-level capability requirements, maps them to representative scenarios and task profiles, then compiles tasks by retrieving or constructing files, dependencies, and runtime setups before generating expert trajectories for SFT. The paper reports 3.6K terminal and 2K 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 reaches 58.4%, roughly matching Claude Opus 4.6 with Claude Code. The resulting data also helps train Nex-N2, an open agent model family that lifts Qwen3.5-35B-A3B to 75.3% on Terminal-Bench 2.1 and 1585 Elo on GDPval.

Original post →

More from coding & agent

coding & agent channel →