The field of large language models has seen rapid advancement, yet how these systems interact, teach, and co-evolve remains underexplored. The paper introduces Environment Duels, a dynamic agentic benchmark that evaluates both capabilities on interactive environments. Each model writes environments, each with a privileged hint, aiming for tasks that capable solvers fail without the hint and pass with it. Across nine frontier LLMs, solving and authoring prove to be distinct skills: the strongest solvers are not the strongest authors, while smaller models learn more from how others played and write the hardest environments and the most useful hints. Beyond the scores, the pairwise interactions form a microcosm of how models affect one another. Head-to-head outcomes are transitive and so yield a stable ranking; independent authors converge on the same designs; and in a reinforcement-learning study with Qwen3.8-27B, the learner's hint lift on an author's environments tracks how much training on those environments improves it on held-out reasoning and agentic benchmarks, whether the environments are its own or another model's. Each of these patterns has a counterpart in how models shape one another outside the duel. By staging these interactions under controlled conditions, Environment Duels offers a window onto how LLMs teach, challenge and learn from one another.
Distinct skills: solving versus authoring
The benchmark simultaneously evaluates two complementary capabilities: the ability of a model to solve tasks within an environment, and the ability to construct environments that challenge other models. The experimental setup shows that these two skills do not necessarily co-occur in the same model. Among nine frontier LLMs examined, top-performing solvers did not necessarily produce the most effective environments. This dissociation suggests that the competencies required for successful task completion differ from those needed to design instructive evaluation suites. Models that excel at following instructions and executing predefined goals may not automatically generate the kind of structured, hint-requiring tasks that probe the limits of other systems.
Smaller models as keen observers and architects
An unexpected finding involves the relative performance of smaller models. While larger models dominate raw solving power, smaller models demonstrated a heightened sensitivity to the patterns introduced by their peers. These models learned more from observing how other agents interacted with their environments and, crucially, wrote the hardest environments and the most useful hints. This result inverts the assumption that environment quality strictly correlates with model size. The data indicate that compact models, trained on more constrained capacities, may develop a more acute awareness of what makes a task both solvable and informative.
Transitive head-to-head outcomes and stable rankings
Pairwise interactions between models produced outcomes that were transitive, meaning if model A typically outperforms model B, and model B outperforms model C, then model A reliably outperforms model C. This transitivity yielded a stable overall ranking across the nine models. The stability of this ranking suggests that the benchmark produces consistent relative evaluations, even as environments are dynamically generated and modified by different authors. Such predictability is valuable for tracking model progress over time and for comparative studies that require a reliable ordering of system capabilities.
Convergent environment design by independent authors
The benchmark also examined what happens when multiple models independently generate environments without direct coordination. Despite the absence of shared instructions or common objectives, independent authors converged on similar designs. This convergence points to underlying regularities in what constitutes a challenging yet solvable task with a privileged hint. The hint itself—information provided to the solver that is withheld from typical evaluation—proved to be a focal point around which environment structure organized itself. Multiple models, starting from different baselines, produced environments where the hint served a comparable role: bridging the gap between task ambiguity and successful resolution.
Reinforcement-learning study: hint lift and transfer
A reinforcement-learning study involving Qwen3.8-27B provided further insight into the transfer dynamics between authoring and solving. The learner was trained on environments authored by various models, then evaluated on held-out reasoning and agentic benchmarks. The key metric was the "hint lift": the degree to which the model's performance on downstream tasks improved after training on a particular author's environments. The results showed that hint lift tracked how much training on those environments improved the learner's performance, regardless of whether the environments originated from the learner itself or from another model. This finding demonstrates that the value of an environment extends beyond the immediate solver-author pair; it generalizes to affect the broader model ecosystem.
Microcosm of model interactions
Beyond the aggregate scores, Environment Duels reveals a microcosm of how LLMs shape one another through the feedback loop of creation and resolution. Models simultaneously act as creators of evaluation suites and solvers of those suites. The observed patterns—distinct solver and author skills, the outsized learning impact on smaller models, transitive head-to-head outcomes, convergent environment design, and hint-mediated transfer—collectively paint a picture of an ecosystem in which each model's output becomes another model's training signal. By staging these interactions under controlled conditions, the benchmark offers a window onto the subtle ways LLMs teach, challenge, and learn from one another.
Code for the benchmark is available at the provided URL, enabling other researchers to replicate the duels or extend the framework to new model families and task domains.