Large language models (LLMs) possess fundamental abilities such as common sense reasoning and coherent language generation. They have been adapted for specific domain tasks such as code generation and mathematical problem solving. This has resulted in exceptional performance of specialized models in specific domains, such as code generation or logical reasoning. However, this raises the question of whether it is possible to combine a general model with a domain-specific reinforcement model to obtain new capabilities. For example, combining a model's code comprehension ability with another model's language generation ability for generating code descriptions. Traditional approaches involve pre-training or fine-tuning the general model on the data used to train the reinforcement model. However, this may require additional computational resources. Using different models can leverage existing capabilities without encountering catastrophic forgetting issues present in traditional methods. To overcome the limitations of training and data mentioned earlier, researchers from Google Research and Google DeepMind propose and explore a practical model combination scenario: (i) access to one or more reinforcement models and a general model, (ii) no changes to the weights of any model, and (iii) access to a limited dataset representing the combined capabilities of the provided models, such as combining code generation and complex logical reasoning. They introduce an innovative framework called Compositional Reinforcement Language Model (CALM) to address the general model combination scenario mentioned above. Unlike simple mixtures of reinforcement and general LMs, CALM introduces a set of trainable parameters in the intermediate layer representations of the reinforcement and general models. CALM aims to find the optimal fusion of these models, improving their overall performance in handling new complex tasks more effectively than any individual model, while preserving each model's unique capabilities. They explore important practical applications of CALM, with a focus on language inclusivity and code generation. In the context of language inclusivity, they use a model specifically trained for low-resource languages. By combining this model with LLM, they endow them with advanced generation and reasoning abilities, significantly improving performance in translation and arithmetic reasoning tasks for low-resource languages. Interestingly, this combined model outperforms the performance of both base models and even outperforms a version of LLM that has undergone further pre-training or LoRA fine-tuning customized for low-resource languages. In the case of code generation, they use a model trained on various open-source code in different programming languages, combining it with LLM. Thus, leveraging its underlying logic and generation abilities, they achieve superior performance in tasks involving code interpretation and completion compared to the two base models.