LatentGym: A Testbed For Cross-Task Experiential Learning With Controllable Latent Structure

A new testbed called LatentGym offers a controlled way to measure whether large language model agents can learn from experience across related tasks, according to a paper posted to arXiv on 13 June 2026 [1]. The suite is built around a ground-truth latent variable that governs the hidden structure shared across tasks, allowing researchers to separate exploration from exploitation [1]. The framework addresses a gap in current evaluation methods, which the authors argue do not provide shared, controllable latent structures and cannot determine whether or why agents improve [1]. LatentGym constructs environments where the underlying task structure is known, yielding metrics that distinguish between an agent's information-gathering actions and its use of that information [1]. This separation is intended to support the study of cross-task experiential learning in settings such as personalization and interactive assistance [1]. The paper presents empirical studies organized around three questions: how and why frontier models fail to adapt across related tasks; whether post-training on related task sequences improves general cross-task adaptation and where those gains originate; and how design choices like inter-task feedback shape training dynamics and generalization [1]. The work builds on a broader machine-learning interest in transfer learning, where models trained on one dataset are fine-tuned on another. A 2022 study on catalyst informatics noted that transfer learning has been used to improve model performance on smaller datasets after initial training on larger ones, and that the small-molecule and drug-discovery communities have seen success transferring between related tasks [3]. LatentGym's construction echoes the principle that controlled, ground-truth environments are necessary to isolate causal factors in agent learning. Without such control, it is difficult to tell whether performance gains come from better exploration strategies or from more effective exploitation of gathered information [1]. The authors demonstrate the suite on sequential, personalized, and interactive settings, aiming to establish a foundation for designing agents that adapt more reliably [1]. The paper was released through arXivLabs, a framework that lets collaborators develop and share new features on the arXiv platform [1]. The research bundle associated with the paper includes code-finding and repository tools such as CatalyzeX and DagsHub, though their specific roles in the work are not detailed in the abstract [2][4].