Environment Duels: A Micro-Ecology of LLMs Teaching and Learning from Each Other

The mutual influence among large models has long gone beyond the level of "who scores higher." Frontier models absorb content produced by previous generations and competitors from synthetic training pipelines and open web data, forming a cycle in which they shape each other. Environment Duels brings this process into a controlled experiment: models both create problems and solve them, allowing us to observe the dual roles they play in the ecology.

Dual Roles: Problem Setter and Problem Solver

Environment Duels is a dynamic agent benchmark that simultaneously evaluates a model's ability to "solve" and to "set" problems in interactive environments. In each round of a duel, the model must write an interactive environment and attach a privileged hint, with the goal of making a capable solver fail without the hint and pass with it.

According to the study, after covering nine frontier models, the two capability dimensions show clear divergence: the strongest solver is not the strongest problem setter. This conclusion itself is not surprising, but it turns "capability fragmentation" from intuition into a quantifiable benchmark result.

Privileged hint is a key mechanism in the Environment Duels design, referring to extra information that the solver normally cannot see and can only obtain under specific conditions. The problem setter needs to embed a hint in the environment such that a sufficiently capable solver fails without this hint but can clear the level with it. The significance of this design is that it treats the hint itself as a quantifiable "teaching signal": whether a hint is effective does not depend on how detailed it is written, but on whether it can truly change the solver's behavioral outcome. Therefore, evaluating the quality of problem setting is transformed into evaluating the "gain" of the hint—that is, the difference in pass rate for the same solver with and without the hint. This mechanism also naturally limits the room for cheating: if the hint is too explicit, all solvers can pass, and an excessively large gain instead indicates that the environment lacks challenge; if the hint has no effect, it indicates that the problem setter failed to identify the real difficulty.

The Counterintuitive Advantage of Small Models

More noteworthy is the mismatch between author capability and model scale. Research shows that smaller models instead learn more from others' duels and can write the hardest environments and the most effective hints.

This suggests a neglected learning path: problem-setting ability relies more on observing "how others fail" than on stacking parameter scale. For training data builders, this means small models can take on the role of generating high-quality environments and hints, rather than merely passively consuming the outputs of large models.

Parameter scale and capability divergence here exhibit a non-monotonic relationship, which is worth understanding from the perspective of training data distribution. Large models usually perform better on a wide range of tasks, but their training objective is biased toward "giving the correct answer" rather than "designing problems that make others fail." Problem-setting ability requires the model to finely model the solver's failure modes, which is closer to a metacognitive ability: understanding where and why others get stuck. Due to capacity constraints, smaller models may instead rely more on inducing from specific failure cases observed in duels, forming a sharper "difficulty localization." In addition, when generating environments and hints, smaller models have a more concentrated output distribution and are more likely to converge to high-information designs, whereas large models may write environments that are too loose or too trivial due to overly broad knowledge coverage. This phenomenon suggests that problem-setting ability may be cultivated separately through targeted fine-tuning or duel-observation data, without waiting for an overall increase in model scale.

Ecological Patterns Behind Duel Results

Beyond scores, the study distills three interaction patterns, each of which has a counterpart in the real model ecology:

• Head-to-head results are transitive, so a stable ranking can be formed;
• Independent authors converge to the same designs, indicating that there is a reproducible structure of "good problems" in environment creation;
• Hint gain predicts transfer effects: in reinforcement learning experiments with Qwen3.8-27B, the degree of hint improvement a learner achieves on a certain author's environment correlates with the degree of improvement on its reasoning and agent benchmarks after training on that environment, regardless of whether the environment comes from itself or other models.

The third is the most practically valuable. It transforms "environment quality" from a subjective evaluation into an observable signal: if a model can greatly improve its score on a certain environment simply by relying on hints, then that environment is likely also effective training material.

Transitivity in the context of games and evaluation refers to the property that if A can consistently beat B and B can consistently beat C, then A can also consistently beat C. Many adversarial evaluations lack transitivity due to cyclic dominance, making it impossible to give a global ranking. Environment Duels observes that head-to-head results are transitive, meaning that duel results can support a consistent strength ordering rather than falling into a "rock-paper-scissors" cycle. This provides the basis for the benchmark's practicality: only with transitivity can scores be interpreted as a monotonic mapping of capability. At the same time, transitivity also implies that there is some comparable latent dimension in the quality of environments and hints, rather than pure dependence on pairing relationships.

Why This Constitutes a "Micro-Ecology"

In the real world, models feed on each other through synthetic data, evaluation problems, code, and public content. The value of Environment Duels lies in compressing these loose, difficult-to-attribute interactions into controllable duels, making "who teaches whom, and how well" measurable.

The authors describe this design as a window for observing how large models teach, challenge, and learn from each other. For teams working on data flywheels and synthetic data pipelines, this benchmark offers an approach: rather than only evaluating how models answer questions, also evaluate how they set them, because problem-setting quality may more directly determine the ceiling of the next round of training.

The code is open-sourced on GitHub (LLM-Duels/EnvDuels) and can reproduce the above duel setup.

Synthetic data flywheel refers to a cyclic pipeline in which data generated by a model is used to train the next generation of models, which then generate more data. The risk of this cycle is that models absorb biases from themselves or their generation, leading to diversity collapse and capability degradation. The perspective offered by Environment Duels is that the quality of the flywheel depends not only on the quality of "answer data" but also on the quality of "problem data." The environments and hints generated by problem setters actually define the distribution of challenges that learners will encounter in the next round of training. If problem setters only produce simple or homogeneous environments, the flywheel will spin idle in low-difficulty regions. Therefore, incorporating problem-setting ability into evaluation is equivalent to adding a "difficulty adjustment valve" to the flywheel, making the richness of training signals an object that can be monitored and optimized.

One-line Summary

Environment Duels lets models both set and solve problems, revealing that problem setting and problem solving are two separable capabilities.

Paragraph Summary

Environment Duels is a dynamic agent benchmark that lets models simultaneously play the roles of problem setter and problem solver in interactive environments: the problem setter must write an environment and attach a privileged hint so that a capable solver fails without the hint and passes with it. After covering nine frontier models, the study finds that the strongest solver is not the strongest problem setter; the two are different skills. Smaller models instead learn more from others' duels and can write the hardest environments and the most effective hints. The study also distills three ecological patterns: head-to-head results are transitive, independent authors converge to the same designs, and hint gain predicts the transfer effect of environment training on reasoning and agent benchmarks. This benchmark compresses the loose mutual influence among models into controllable duels, providing a measurable approach for synthetic data flywheels and problem-setting ability evaluation. The code is open-sourced.