Co-evolution with GenEnv Accelerates AI Learning on Dynamic, Five-Benchmark Tasks

The challenge of training truly capable artificial intelligence agents currently faces a significant bottleneck, stemming from the expense and limitations of real-world training data. To address this, Jiacheng Guo, Ling Yang, Peter Chen, and colleagues introduce GenEnv, a novel framework that establishes a dynamic co-evolution between an agent and a generative environment simulator. Unlike conventional methods relying on fixed datasets, GenEnv creates a continuously evolving curriculum, with the simulator generating tasks specifically matched to the agent’s current skill level, a process guided by a carefully designed reward system. Evaluations across five diverse benchmarks, including API-Bank and TravelPlanner, demonstrate that GenEnv improves agent performance by over 40% compared to existing models, and achieves comparable results to much larger systems while using significantly less data, offering a data-efficient pathway towards more powerful and adaptable AI agents.

Dynamic Environments Boost Language Agent Learning

The research team developed GenEnv, a novel framework that addresses limitations in training large language model (LLM) agents by overcoming the high cost and static nature of real-world interaction data. GenEnv establishes a co-evolutionary game between an agent and a scalable, generative environment simulator, creating a dynamic curriculum tailored to the agent’s capabilities. This process is guided by a “Curriculum Reward” that aligns task difficulty with the agent’s current skill level, fostering continuous learning and improvement. Experiments across five benchmarks, API-Bank, ALFWorld, BFCL, Bamboogle, and TravelPlanner, demonstrate that GenEnv significantly enhances agent performance.

Specifically, the team achieved up to a +40.3% improvement on ALFWorld compared to 7B baseline models, and the system matches or exceeds the performance of larger models on this suite of tasks. The data reveals that GenEnv achieves better performance than offline data augmentation using Gemini 2.5 Pro while utilizing 3.3times less data, highlighting its data efficiency.

Further analysis shows that GenEnv induces an emergent curriculum, where the environment simulator gradually increases task complexity, as evidenced by a rise in task description length. Correspondingly, the agent’s reasoning chains, measured by response length, increase over time, and the agent’s success rate remains stable within a controlled band despite the rising difficulty. The team recorded an increase in average agent response length from 137 tokens to 204 tokens across six epochs, demonstrating the agent’s growing capacity for complex reasoning. The research team also investigated the impact of different data generation strategies, comparing GenEnv to variants that use random or static data generation. Results demonstrate that aligning the simulator with an intermediate success rate maximizes learning signal for the agent and yields a stable, difficulty-calibrated curriculum. The team’s work delivers a data-efficient pathway for scaling agent capabilities, offering a significant advancement in the field of LLM agent training.

Dynamic Environments Accelerate Agent Learning

GenEnv represents a significant advance in training artificial intelligence agents, shifting the focus from static datasets to a dynamic, co-evolutionary game between the agent and a generative environment simulator. This framework establishes a continuous loop where the simulator creates tasks tailored to the agent’s current capabilities, effectively providing a curriculum aligned with the agent’s zone of proximal development. Evaluations across five challenging benchmarks demonstrate that GenEnv improves agent performance by up to 40.3% compared to existing methods and matches or exceeds the performance of larger models.

Importantly, the system achieves these results while using considerably less data than alternative approaches, such as those based on offline data augmentation. The research highlights the value of faithful and diverse environment simulation in creating high-quality training data and benchmarks. By casting environment design as a learnable policy with its own reward signal, GenEnv offers a data-efficient pathway for scaling agent capabilities. The authors acknowledge that the current implementation is specific to the chosen tasks and models, and future work could explore broader applications and alternative simulator designs. They suggest that difficulty-aligned simulators hold promise as a general recipe for training robust agents in domains where real-world exploration is costly or risky, paving the way for more adaptable and efficient artificial intelligence systems.

Dynamic Environments Boost Language Agent Learning

The research team developed GenEnv, a novel framework that addresses limitations in training large language model (LLM) agents by overcoming the high cost and static nature of real-world interaction data. GenEnv establishes a co-evolutionary game between an agent and a scalable, generative environment simulator, creating a dynamic curriculum tailored to the agent’s capabilities. This process is guided by a “Curriculum Reward” that aligns task difficulty with the agent’s current skill level, fostering continuous learning and improvement. Experiments across five benchmarks, API-Bank, ALFWorld, BFCL, Bamboogle, and TravelPlanner, demonstrate that GenEnv significantly enhances agent performance.

Specifically, the team achieved up to a +40.3% improvement on ALFWorld compared to 7B baseline models, and the system matches or exceeds the performance of larger models on this suite of tasks. The data reveals that GenEnv achieves better performance than offline data augmentation using Gemini 2.5 Pro while utilizing 3.3times less data, highlighting its data efficiency.

Further analysis shows that GenEnv induces an emergent curriculum, where the environment simulator gradually increases task complexity, as evidenced by a rise in task description length. Correspondingly, the agent’s reasoning chains, measured by response length, increase over time, and the agent’s success rate remains stable within a controlled band despite the rising difficulty. The team recorded an increase in average agent response length from 137 tokens to 204 tokens across six epochs, demonstrating the agent’s growing capacity for complex reasoning. The research team also investigated the impact of different data generation strategies, comparing GenEnv to variants that use random or static data generation. Results demonstrate that aligning the simulator with an intermediate success rate maximizes learning signal for the agent and yields a stable, difficulty-calibrated curriculum. The team’s work delivers a data-efficient pathway for scaling agent capabilities, offering a significant advancement in the field of LLM agent training.