All Agents Are Evolving Translators

Preface

This is my very initial draft of the research proposal. I know there’s a lot of work to do (and, of course, many modifications), so I am posting here to gather more feedback. I welcome all collaboration opportunities. For instance, what is the branch of mathematics that property 4.1 belongs to? What are those papers share similar ideas with mine (module 1, module 2, and module 3)? What may other research be conducive to accomplishing these goals mentioned in this research proposal?

Introduction

To build a GUI Agent able to adapt to users’ needs or an Automated Trading System, it’s important to construct an architecture that is fast, interpretable, of high performance, and industry-oriented. Despite the prominent achievements of deep learning in Large Language Models (LLMs), the interpretability of neural network architectures is still poor, which affects their credibility and hence limits the deployments of risk-sensitive fields. In certain scenario-specific domains with scarce data, rapidly obtaining a large number of supervised learning labels is challenging, and the workload of manually labeling data would be enormous. Catastrophic forgetting in neural networks further leads to low data utilization rates. In situations where swift responses are vital, the density of the model makes local deployment difficult and the response time long, which is not conducive to local applications of these fields. To tackle these problems, this research is designed to construct a lightweight and interpretable architecture consisting of these 3 parts: encoder or decoder, decision-making algorithm, and continuous self-evolving mechanism.

This research will propose an architecture composed of 3 modules. The first is the perception module, Parallel Nested State Spaces Framework Ensemble, which takes in an orchestration chart and outputs a processed orchestration chart. Each output track matches with 2 propositions’ states (one and its conjugate) in module 2, the decision-making module, Believability-Weighted Expert Teams Algorithm. These 2 modules together form the base of the architecture, which influences agents’ performance directly (just like the neural network itself). Then, an independent continuous self-evolving mechanism, the Independent Agent Learning Algorithm, will train the aforementioned architecture (like the backpropagation algorithm).