# ProgramGym builds 1,648 verifiable training environments for coding agents

> The October 9 arXiv preprint describes environments for reproducing programs from scratch and a staged agent workflow.

- Canonical URL: https://newsverum.com/en/news/programgym-training-environments-coding-agents-arxiv-2026
- Publisher: Newsverum
- Section: Technology
- Country: China
- Published: 2026-10-13T10:04:10+00:00
- Updated: 2026-10-13T10:04:10+00:00
- Event / data period: 2026-10-09
- Other editions: [Türkçe](https://newsverum.com/tr/news/programgym-training-environments-coding-agents-arxiv-2026.md), [العربية](https://newsverum.com/ar/news/programgym-training-environments-coding-agents-arxiv-2026.md), [Français](https://newsverum.com/fr/news/programgym-training-environments-coding-agents-arxiv-2026.md), [Español](https://newsverum.com/es/news/programgym-training-environments-coding-agents-arxiv-2026.md), [Deutsch](https://newsverum.com/de/news/programgym-training-environments-coding-agents-arxiv-2026.md), [Português](https://newsverum.com/pt/news/programgym-training-environments-coding-agents-arxiv-2026.md)

Researchers from Xiamen University and other institutions introduced ProgramGym, a framework for training coding agents to inspect a program’s behavior and reproduce it from scratch. In the preprint submitted to arXiv on October 9, Program-Env creates training environments with executable reference programs, isolated sandboxes, usage documentation and black-box tests. The authors report building 1,648 environments in C, C++, Go and Rust. Program-Agent then separates analysis, implementation and review into distinct roles, with compilation checks and recovery checkpoints between iterations.

On ProgramBench, the authors report that fine-tuning Qwen3.8-27B on trajectories from the generated environments raised its test-pass result from 34.87% to 42.84%. With Program-Agent, the base model scored 55.71% and the fine-tuned version 56.25%. The paper defines test pass as a macro-average of task-level pass rates, so these figures do not mean every task was solved. The results concern the authors’ program-reproduction tasks and stated tool setup; they are not a measure of general coding ability.

## Data

### ProgramGym training environment count

Program-reproduction environments reported in the paper.

Methodology: Authors count environments built for C, C++, Go and Rust.

| Measure | Value |
|---|---|
| Program-Env training environments | 1648 |

Sources: <https://arxiv.org/html/2610.13395v1> · Download data: [JSON](https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/training-environment-count?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/training-environment-count?lang=en&format=csv)

### r

Test-pass rates reported for the authors’ model and agent-scaffold pairings.

Methodology: ProgramBench defines this metric as the macro-average of task-level test-pass rates. The Program-Env comparison holds Claude Code fixed; the Program-Agent comparison pairs the same model with Claude Code or Program-Agent. The tasks are not thereby fully solved.

| Model and harness | Tests Passed (%) |
|---|---|
| Qwen3.8-27B + Claude Code, before fine-tuning on Program-Env trajectories | 34.87 |
| Qwen3.8-27B-SFT + Claude Code, after fine-tuning on Program-Env trajectories | 42.84 |
| Qwen3.8-27B + Program-Agent, same base model | 55.71 |
| Qwen3.8-27B-SFT + Program-Agent, same fine-tuned model | 56.25 |

Sources: <https://arxiv.org/html/2610.13395v1> · Download data: [JSON](https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/programbench-test-pass?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/programbench-test-pass?lang=en&format=csv)

## Sources

- [arXiv preprint by researchers affiliated with Xiamen University, Washington University in St. Louis, the Chinese Academy of Sciences and Unisound](<https://arxiv.org/html/2610.13395v1>) (Source checked: 2026-10-13; Published: 2026-10-09)

---

Cite the canonical URL above. [Editorial standards and corrections](https://newsverum.com/en/editorial)