‹ back to 2026-08-17

Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL

arXiv cs.CL LLM Agents Xiaojun Wu, Cehao Yang, Honghao Liu, Xueyuan Lin, Zhichao Shi, Hao Zhou, Xuhui Jiang, Chengjin Xu, Jia Li, Jian Guo 2026-08-14

TL;DR - Envs-FORGE synthesizes difficulty-targeted, reward-verified environments for training terminal agents. It improves agent performance over fixed task-generation recipes while holding training-set size and synthesis scale roughly constant.

Uses estimated seed pass rates and a per-seed MILP to select among six environment-synthesis actions near the learning frontier.
Jointly rewrites instructions, fixtures, oracle solutions, tests, and Docker environments; only gold-verified bundles enter RL training.
On Qwen 3.5 35B, it raises Pass@1 by 9.2 points on tb-core and 6.4 points on tb-2.0 versus Base.
It reaches 77.1% on SWE-bench Verified versus 73.4% for Base and shows gains across evaluated 4B–35B models.

view merged work →