Retrieval-Augmented Thought Process as Sequential Decision Making

Original Paper: https://arxiv.org/abs/2402.07812

By: Thomas Pouplin, Hao Sun, Samuel Holt, Mihaela van der Schaar

Abstract:

Large Language Models (LLMs) have demonstrated their strong ability to assist people and show "sparks of intelligence". However, several open challenges hinder their wider application: such as concerns over privacy, tendencies to produce hallucinations, and difficulties in handling long contexts. In this work, we address those challenges by introducing the Retrieval-Augmented Thought Process (RATP). Given access to external knowledge, RATP formulates the thought generation of LLMs as a multiple-step decision process. To optimize such a thought process, RATP leverages Monte-Carlo Tree Search, and learns a Q-value estimator that permits cost-efficient inference. In addressing the task of question-answering with private data, where ethical and security concerns limit LLM training methods, RATP achieves a 50% improvement over existing in-context retrieval-augmented language models.

Summary Notes

Enhancing Language Models: The Power of Retrieval-Augmented Thought Process

Language models have significantly advanced, understanding and generating text like humans. Yet, they struggle with detailed, sensitive data and often make mistakes.

A promising solution is the Retrieval-Augmented Thought Process (RATP), which boosts language models by adding external knowledge, improving their output on complex tasks.

What is RATP?

RATP transforms language models by:

• Thinking in Sequences: It views generating thoughts as a series of decisions, allowing for a logical integration of various knowledge sources.
• Using Monte-Carlo Tree Search (MCTS): This technique helps RATP efficiently sort through and integrate knowledge.
• Applying a Q-value Estimator: This ensures each thought step is relevant and impactful.
• Enhancing Complex Task Performance: Demonstrated improvements on tasks like BoolQA and emrQA show RATP's ability to boost language model capabilities.

Breaking Down Thought Generation

RATP sees thought generation as a Markov Decision Process (MDP), involving:

• States and Actions: States are previous thoughts and actions, and the action space can include external documents or past thoughts.
• Transition Dynamics: Combining these elements generates new thoughts, mimicking human thought processes.
• Reward Function: The accuracy of answers helps refine the process for better results.

The Role of Monte Carlo Tree Search

MCTS is crucial for RATP, given its complex decision-making needs, and involves:

• Selection and Expansion: Choosing which thought to develop further and integrating new information to generate thoughts.
• Simulation and Backpropagation: Evaluating new thoughts and updating the decision tree for continuous improvement.

Innovative Scoring Models

RATP uses two scoring models to value thoughts:

• Offline Model-Based Estimation: Predicts the value of new thoughts using past data.
• Self-Critic Method: Allows the language model to evaluate its outputs for more accurate assessments.

Experiments and Results

Testing RATP has shown:

• Better Handling of Sensitive Information: A 50% improvement in private knowledge scenarios.
• Superior Performance on Boolq Dataset: Demonstrating RATP’s advanced external knowledge integration and thought optimization.

Conclusion

RATP enhances language models by integrating external knowledge and treating thought generation as a decision-making process.

With MCTS and innovative scoring models, RATP overcomes current limitations, offering a path to more versatile and efficient language models.

Impact Statement

RATP’s benefits extend to making advanced language model capabilities more accessible and cost-effective, especially for dealing with sensitive data.

Its documentation of the thought process also enhances interpretability and accountability in AI decision-making, marking progress towards more reliable and transparent AI systems.