BWArea Model: Learning World Model, Inverse Dynamics, and Policy for Controllable Language Generation

TL;DR

The BWArea model introduces a brain-inspired decision-making framework for controllable language generation, combining a world model, inverse dynamics, and policy to improve performance and alignment with downstream tasks.

Contribution

It proposes a novel decision-making based architecture for language models, inspired by human brain mechanisms, enhancing controllability and robustness against data noise.

Findings

Achieves competitive performance with 1B parameter LLMs using 30B tokens.

Maintains performance with dirty data, unlike auto-regressive models.

Outperforms auto-regressive LLMs on 9 out of 10 tasks in benchmarks.

Abstract

Large language models (LLMs) have catalyzed a paradigm shift in natural language processing, yet their limited controllability poses a significant challenge for downstream applications. We aim to address this by drawing inspiration from the neural mechanisms of the human brain, specifically Broca's and Wernicke's areas, which are crucial for language generation and comprehension, respectively. In particular, Broca's area receives cognitive decision signals from Wernicke's area, treating the language generation as an intricate decision-making process, which differs from the fully auto-regressive language generation of existing LLMs. In a similar vein, our proposed system, the BWArea model, conceptualizes language generation as a decision-making task. This model has three components: a language world model, an inverse dynamics model, and a cognitive policy. Like Wernicke's area, the inverse dynamics model is designed to deduce the underlying cognitive intentions, or latent actions, behind each token. The BWArea model is amenable to both pre-training and fine-tuning like existing LLMs. With 30B clean pre-training tokens, we have trained a BWArea model, which achieves competitive performance with LLMs of equal size (1B parameters). Unlike fully auto-regressive LLMs, its pre-training performance does not degenerate if dirty data unintentionally appears. This shows the advantage of a decomposed structure of BWArea model in reducing efforts in laborious data selection and labeling. Finally, we reveal that the BWArea model offers enhanced controllability via fine-tuning the cognitive policy with downstream reward metrics, thereby facilitating alignment with greater simplicity. On 9 out of 10 tasks from two suites, TextWorld and BigBench Hard, our method shows superior performance to auto-regressive LLMs.