In plain words
Environment Duels is a test in which nine language models both write interactive tasks and try to solve tasks written by other models. Each task comes with a hint from its author.
Researchers compare how often models solve tasks without hints, how challenging each author's tasks are, and whether the hints change success.
The best task solvers were not the best task authors. In a separate training comparison, one model improved more on five tests after training on tasks from an author whose hints had helped it more. That result applies to the learner and authors tested.
The research
If models learn from tasks other models create, how well those tasks are written may shape what later models learn. The paper asks whether models that solve tasks well can also create useful challenges and guidance for others.
Environment Duels has nine language models write interactive tasks with hints, solve one another’s tasks, and measures solving and task-authoring separately.
The paper asks whether models can create useful challenges for other models to learn from, as well as solve tasks themselves. Environment Duels tests both roles by having nine models write interactive tasks with hints and solve one another’s tasks, measuring solving and authoring separately.
The strongest no-hint solvers were not the strongest authors. In a separate training comparison, Qwen3.8-27B gained more on five tests from an author's environments when its pre-training hint lift on those environments was higher. This supports hint lift as a signal of training usefulness in that comparison.
The training comparison covers one learner, Qwen3.8-27B, trained on environments from four authors and evaluated on five tests. It suggests hint lift can signal training usefulness in this setup, but does not establish that relationship for other learners or authors.
Analysis
Deep
- Original paper authors
- Shibo Jin, Zhiqiu Xu, Yiling Zou, Shangyu Gong, Mayur Naik
- Submitted to arXiv
- License
- CC BY 4.0