Inside CALM: Google DeepMind’s Unique Method to Augment LLMs with Other LLMs
CALM provides a clever approach for composing LLMs specialized on different tasks.
I recently started an AI-focused educational newsletter, that already has over 160,000 subscribers. TheSequence is a no-BS (meaning no hype, no news, etc) ML-oriented newsletter that takes 5 minutes to read. The goal is to keep you up to date with machine learning projects, research papers, and concepts. Please give it a try by subscribing below:
TheSequence | Jesus Rodriguez | Substack
The best source to stay up-to-date with the developments in the machine learning, artificial intelligence, and data…
Knowledge augmentation is one of most important topics in LLM-based applications. Over the last few months, we have seen a proliferation of augmentation techniques such as retrieve-augmented-generation(RAG) that attempt to expand LLM knowledge with access to external tools or data. However, can we augment LLMs with other LLMs? This seems like an area worth exploring and is the subject of a new paper by Google DeepMind.
The idea of augmenting LLMs with LLMs ties directly into the area of model composition. The key principle to explore is whether it’s possible to combine a general-purpose anchor model with a specialized model to create new abilities. For example, could the code understanding ability of one model be merged with the language generation skill of another to facilitate code-to-text generation? Typically, the solution involves additional training or fine-tuning of the anchor model using the data from the specialized model. However, this approach can be computationally expensive and sometimes impractical due to data privacy concerns and organizational limits.
To overcome these challenges, Google DeepMind has proposed a new method called Composition to Augment Language Models (CALM) for model composition. This approach does not alter the fundamental structure of the models. Instead, it involves working with an anchor model and one or more augmenting models without modifying their core algorithms. Additionally, this method only requires a minimal amount of data that represents the combined capabilities of the models involved, such as the integration of code generation with advanced logical reasoning. This innovative approach not only conserves resources but also maintains the integrity and individual strengths of each model.
The CALM Architecture
CALM introduces a very clever architecture for composing LLMs in highly effective way. Unlike simpler methods of combining models, CALM introduces a minimal number of trainable parameters to work with the intermediate layers of both the anchor and augmenting models. This method allows for a more effective integration, enabling the performance of new, complex tasks that neither model could achieve independently. Importantly, this process preserves the individual strengths of each model.
CALM aims to expand the capabilities of a primary LLM by integrating it with specialized augmenting models, each with unique abilities. For example, it can work with models that specialize in key-value mapping, low-resource languages, or coding. During this process, the core structures of these models remain unchanged. CALM simply adds a few parameters that learn from the layer representations of the models. This approach has shown remarkable results, such as enabling arithmetic operations on key-value pairs, a task beyond the reach of either model alone.
CALM focuses on the interaction between a chosen set of layers of two target models. Their methodology involves the introduction of two specific sets of additional parameters to these layers. The first set consists of linear transformations. These transformations are responsible for adapting the layer representations from the first to match the dimensionality of the representations in the second model. This ensures that the data from both models is compatible, facilitating a smoother integration of the two models.
The second set comprises cross-attention layers. These layers are designed to create a dynamic interaction between the models. By implementing these cross-attention layers, CALM enables the models to effectively share and process information, enhancing their combined output. This sophisticated integration through linear transformations and cross-attention layers is a key component of the CALM framework, allowing for an efficient and harmonious blend of the distinct capabilities of each model.
The CALM framework is particularly effective in scenarios where there’s a need to leverage specialized knowledge stored in different models. For instance, a foundational LLM can be augmented with models containing proprietary data or expertise, enhancing its capabilities in areas like reasoning, world knowledge, and language generation in specific domains. What makes CALM stand out is its ability to merge distinct knowledge from multiple models without the need to update the individual models. This integration is achieved through a few trainable cross-attention parameters. Google DeepMind’s experiments with CALM have consistently shown that it effectively utilizes the strengths of both models, leading to significant improvements in the anchor model’s performance in various challenging tasks, including translation for low-resource languages, reasoning, and code explanation and generation.
Google DeepMind evaluated CALM across different industry benchmark for tasks such as coding, math and traditional language tasks. Most of the evaluations included started with a baseline PALM model and evolving into a CALM architecture.
The following table shows the results in math benchmarks.
Similarly, we can find the results for code interpretation and generation.
Composing LLMs to augment the knowledge of a target model is a very intuitive and clever idea. As more knowledge becomes compacted into LLMs, architectures such as CALM are likely to become more relevant. Google DeepMind’s can definitely inspire others to push boundaries in this area.