Environment Duels: A Microcosm of LLM Co-Evolution
The strongest problem-solvers were not the strongest task authors in a benchmark of nine frontier LLMs.
Environment Duels asks models to build interactive tasks that capable solvers fail without a privileged hint but pass with it. Smaller models learned more from others’ play and produced the hardest environments and most useful hints. In a reinforcement-learning study with Qwen3.8-27B, the improvement from hints tracked how much training on those environments improved performance on held-out reasoning and agentic benchmarks. That relationship held for environments written by the learner itself or by another model.
ArXiv · AI/CL/LG's note
Environment Duels asks models to build interactive tasks that capable solvers fail without a privileged hint but pass with it. Smaller models learned more from others’ play and produced the hardest environments and most useful hints. In a reinforcement-learning study with Qwen3.8-27B, the improvement from hints tracked how much training on those environments improved performance on held-out reasoning and agentic benchmarks. That relationship held for environments written by the learner itself or by another model.
ArXiv · AI/CL/LG's note
score 5