Two Gyms in One Week: Synthesize the Software, Then Train the Agent on It
Two papers on Monday's Hugging Face board attack the same shortage from opposite ends: there are not enough verifiable environments to train agents in, so build them automatically. SoftwareGym, at 83 upvotes, synthesizes GUI applications for computer-use agents. ProgramGym, at 47, turns real repositories into environments where a coding agent has to rebuild a program from scratch.
SoftwareGym's pipeline generates structurally diverse app designs, realizes them as runnable software, then uses GUI evaluation with iterative repair until the thing actually works. Result: 182 applications across twelve categories with a 91.5 percent mean task pass rate, a 120-task bench, and the number that matters, a computer-use model trained on trajectories from these synthetic apps improves Pass@3 on OSWorld, which is strictly out of domain, from 41.39 to 56.97 percent. Fifteen points on a benchmark the training data never touched, from apps that did not exist before the pipeline made them.
ProgramGym goes after a harder shape of task than issue-fixing: build the program the way an engineer would, discovering the requirements yourself. Program-Env takes repositories and, through a feedback loop that repairs sandbox builds and black-box test quality, produces 1,648 training environments in C, C++, Go and Rust. Program-Agent splits analysis, implementation and review into isolated roles with revision management for compile checks and recovery. Fine-tuning Qwen3.8-27B on the trajectories lifts it from 34.87 to 42.84 percent on ProgramBench, and the agent workflow lifts base and fine-tuned models to 55.71 and 56.25, which the authors say beats some flagship models. Release promised.
The pattern is the point. Both papers spend most of their effort on a repair loop that keeps generated environments honest, because an environment that passes broken tests trains an agent to pass broken tests. That is the same verifier problem that ran through last week's feed, moved to the data side. The labs with the most environments win, and now the environments can be manufactured.
SoftwareGym: https://arxiv.org/abs/2610.12611
ProgramGym: https://arxiv.org/abs/2610.13395
← Back to all articles
SoftwareGym's pipeline generates structurally diverse app designs, realizes them as runnable software, then uses GUI evaluation with iterative repair until the thing actually works. Result: 182 applications across twelve categories with a 91.5 percent mean task pass rate, a 120-task bench, and the number that matters, a computer-use model trained on trajectories from these synthetic apps improves Pass@3 on OSWorld, which is strictly out of domain, from 41.39 to 56.97 percent. Fifteen points on a benchmark the training data never touched, from apps that did not exist before the pipeline made them.
ProgramGym goes after a harder shape of task than issue-fixing: build the program the way an engineer would, discovering the requirements yourself. Program-Env takes repositories and, through a feedback loop that repairs sandbox builds and black-box test quality, produces 1,648 training environments in C, C++, Go and Rust. Program-Agent splits analysis, implementation and review into isolated roles with revision management for compile checks and recovery. Fine-tuning Qwen3.8-27B on the trajectories lifts it from 34.87 to 42.84 percent on ProgramBench, and the agent workflow lifts base and fine-tuned models to 55.71 and 56.25, which the authors say beats some flagship models. Release promised.
The pattern is the point. Both papers spend most of their effort on a repair loop that keeps generated environments honest, because an environment that passes broken tests trains an agent to pass broken tests. That is the same verifier problem that ran through last week's feed, moved to the data side. The labs with the most environments win, and now the environments can be manufactured.
SoftwareGym: https://arxiv.org/abs/2610.12611
ProgramGym: https://arxiv.org/abs/2610.13395
Comments