EAGER is a novel generative recommender system that integrates behavioral and semantic information for improved recommendation. The system uses a two-stream generation architecture with a shared encoder and two separate decoders to decode behavior and semantic tokens. It also includes a global contrastive task with summary tokens to enhance discriminative decoding and a semantic-guided transfer task to promote cross-interactions between behavior and semantics. The model is evaluated on four public benchmarks, demonstrating superior performance compared to existing methods. The framework addresses three key challenges: (1) a unified generative architecture for handling two types of information, (2) sufficient and independent learning for each type, and (3) subtle interactions that enhance collaborative information utilization. The model's effectiveness is validated through extensive experiments, showing that it outperforms traditional and generative methods in terms of recall and NDCG metrics. The system is designed to be efficient and robust, with a confidence-based ranking strategy for merging results from different streams. The paper also discusses the impact of various hyperparameters and ablation studies, highlighting the importance of the two-stream generation architecture, global contrastive task, and semantic-guided transfer task in achieving the model's performance. The results demonstrate that EAGER is a promising approach for generative recommendation, combining behavior and semantics for better item recommendations.EAGER is a novel generative recommender system that integrates behavioral and semantic information for improved recommendation. The system uses a two-stream generation architecture with a shared encoder and two separate decoders to decode behavior and semantic tokens. It also includes a global contrastive task with summary tokens to enhance discriminative decoding and a semantic-guided transfer task to promote cross-interactions between behavior and semantics. The model is evaluated on four public benchmarks, demonstrating superior performance compared to existing methods. The framework addresses three key challenges: (1) a unified generative architecture for handling two types of information, (2) sufficient and independent learning for each type, and (3) subtle interactions that enhance collaborative information utilization. The model's effectiveness is validated through extensive experiments, showing that it outperforms traditional and generative methods in terms of recall and NDCG metrics. The system is designed to be efficient and robust, with a confidence-based ranking strategy for merging results from different streams. The paper also discusses the impact of various hyperparameters and ablation studies, highlighting the importance of the two-stream generation architecture, global contrastive task, and semantic-guided transfer task in achieving the model's performance. The results demonstrate that EAGER is a promising approach for generative recommendation, combining behavior and semantics for better item recommendations.