TLDR: HdLM is a new language model architecture that enables hierarchical thinking by allowing different internal layers to decode text simultaneously. This adaptation of existing LLMs improves performance on hierarchical tasks like classification and generation, and offers computational efficiencies, paving the way for more structured AI reasoning.
Large language models (LLMs) like GPT and LLaMA have shown incredible abilities in understanding and generating human language. However, they typically process information and generate responses in a linear fashion, decoding only from their final layer. This approach can sometimes fall short when dealing with complex tasks that require a more structured, step-by-step, or hierarchical way of thinking, similar to how humans approach problems.
Inspired by human hierarchical thinking, researchers have introduced a new approach called the Hierarchical decoding Language model, or HdLM. This innovative model aims to give LLMs the ability to think and generate text at multiple levels of abstraction simultaneously, rather than just from the very end of their processing pipeline.
How does HdLM achieve this? Instead of building a new model from scratch, the researchers adapted existing powerful language models. They essentially copied the “language heads” (the parts responsible for generating text) from the model’s last layer to several selected intermediate layers. These newly added heads are then fine-tuned with different types of task inputs, allowing each selected layer to learn to decode meaningful content at different hierarchical levels.
For example, an earlier layer might generate a coarse-grained, strategic decision, while a later layer refines this into a detailed, fine-grained response. This allows the model to explicitly plan immediate steps and guide subsequent generations, leading to more coherent and logically structured outputs.
The benefits of HdLM are significant. Through extensive experiments, it has been validated that these intermediate layers can indeed produce sensible and relevant content. More importantly, HdLM has achieved state-of-the-art performance across various complex tasks. These include hierarchical text classification (categorizing text from broad to specific labels), classification-guided generation (where a classification result informs the subsequent text generation), and hierarchical text generation (like generating a thought process before providing a final answer).
HdLM has outperformed existing baselines on several datasets, including WoS, DBpedia, ESconv, and EmpatheticDialogues, as well as various cognitive tests. Beyond its improved performance, the research also highlights computational savings during both training and inference, making it a more efficient solution. This study opens up exciting possibilities for developing more generalized hierarchical reasoners in artificial intelligence, potentially even leading to models pretrained from scratch with this inherent hierarchical capability.
Also Read:
- Dynamic Tree Reasoning with Reinforcement Learning for Adaptive LLM Problem Solving
- Early Warning for AI Safety: Monitoring Language Model Thoughts
For more in-depth information, you can read the full research paper here: Making Language Model a Hierarchical Classifier and Generator.
