Two of the most upvoted papers on Hugging Face's daily papers page for October 12, 2026 make the same bet. The next gains for agents will come from generating training environments at scale, not from more hand-built benchmarks. SoftwareGym (arXiv 2610.12611, posted October 8) does it for computer-use agents. ProgramGym (arXiv 2610.13395, posted October 9) does it for coding agents that build whole programs.
SoftwareGym: synthetic desktop software
The Shanghai Jiao Tong University team argues that real native apps take a lot of effort to set up and risk overfitting to a few applications, while browser-hosted environments are too shallow. SoftwareGym is a pipeline that designs GUI applications with varied interfaces, builds them as runnable software, and then uses GUI evaluation with iterative repair until they work. The authors report 182 applications across twelve categories, with a mean GUI task pass rate of 91.5%, as their reliability check.
The result that matters is transfer. A model trained on trajectories from these synthetic apps improved Pass@3 on OSWorld, which the authors call strictly out-of-domain, from 41.39% to 56.97%. It scored 40.0% on their own 120-task SoftwareGym-Bench. Code is on GitHub.
ProgramGym: rebuild the program, don't just patch it
ProgramGym targets a harder coding setting than fixing an issue in an existing repository. The agent must explore a reference program, work out what it needs to do, and design and build it. Its Program-Env component turns repositories into verifiable environments through a feedback loop that repairs sandbox builds and improves black-box tests. The authors report 1,648 environments spanning C, C++, Go and Rust.
Fine-tuning Qwen3.8-27B on trajectories from these environments raised its ProgramBench score from 34.87% to 42.84%. Wrapping models in Program-Agent, a role-separated analyse/implement/review workflow with revision management, lifted the base and fine-tuned models to 55.71% and 56.25%. The authors say that exceeds "some flagship models". For scale, Meta's Muse Wonder model card, released the same day, reports 59% for Wonder Full and 70.5% for Claude Opus 5.5 on ProgramBench.
Why the pairing matters
Both papers point to the bottleneck Karpathy described in a widely shared post the day before: capability is strongest where rewards are verifiable. Each paper is a recipe for making more domains verifiable. Generate the environment, test that it actually works, then train against it. And both find that the harness matters almost as much as the weights. Program-Agent's workflow added more to ProgramBench than fine-tuning did.
What remains uncertain. Both are preprints, not yet peer reviewed. ProgramGym says it "will release" its environments, so they are not yet public. Synthetic environments can also teach the quirks of their generator, and OSWorld transfer is one benchmark, not proof of general robustness.
Analysis: For teams training or fine-tuning agents in-house, the practical lesson is to put engineering effort into the environment factory and its verifier, not only the model. A repair loop that rejects broken environments is what turns cheap synthetic data into supervision worth training on.
SoftwareGym (182 synthesized desktop apps) lifted a computer-use agent's out-of-domain OSWorld Pass@3 from 41.39% to 56.97%, and ProgramGym (1,648 repo-derived environments) lifted Qwen3.8-27B on ProgramBench from 34.87% to 42.84%, with its agent workflow adding more on top; both are preprints.