LatentGym: New Testbed for Cross-Task Experiential Learning in AI Agents

This paper introduces LatentGym, a new testbed specifically designed to facilitate research into continually learning agentic systems. The core vision is to enable agents to become more effective over time by inferring shared hidden structures as they encounter sequences of related tasks, thereby improving future decision-making. Existing training and evaluation frameworks often lack shared, controllable latent structures and adequate metrics to assess how and why agents improve in cross-task experiential learning. LatentGym addresses this by organizing each environment around a ground-truth latent variable that governs the structure across tasks. The testbed provides distinct metrics to differentiate between an agent's exploration capabilities (gathering information about the latent structure) and its exploitation capabilities (using gathered information). Empirical studies using LatentGym have explored why frontier models struggle with cross-task adaptation, how post-training on related tasks improves generalization, and the impact of design choices like inter-task feedback on training dynamics. This work establishes a controlled foundation for advancing LLM agents in sequential, personalized, and interactive settings.