The abstract opens with a pointed observation: language models are usually evaluated in isolation, yet they are increasingly deployed as parts of larger systems—routed between, made to debate, ensembled, or collaboratively trained. For a model inside such a system, the authors argue, what matters most is not how well it performs on its own, but how it contributes to the overall outcome.
To address this gap, the paper introduces "Compositional RL," a framework for evaluating and training LLMs specifically for collaborative reasoning. The idea appears to be that reinforcement learning can be applied at the level of the composition—optimizing a model's behavior in the context of other models—rather than only on individual tasks.
The abstract is brief and does not include experimental details or results, so the concrete methods and findings remain unspecified. Still, the motivation is clear: as LLM deployments become more multi-agent, the standard practice of solo evaluation may be increasingly out of step with real-world use.