OlaGPT boosts LLMs with human-like cognitive modules and intelligent components

OlaGPT is a new framework that aims to enhance the problem-solving abilities of large language models (LLMs) by simulating the human way of thinking. This model incorporates diverse cognitive modules and intelligent mechanisms, such as attention, memory, learning, reasoning, action selection, and decision-making.

In other words, OlaGPT tries to imitate how the human brain performs tasks like paying attention, remembering, learning, thinking, choosing and making decisions.

OlaGPT was evaluated on various datasets that measure reasoning abilities and proved its superiority over existing language models benchmarks.

Current LLMs use specific prompts to generate chains of thought (CoT). Different prompt engineering methods are used to improve their performance including CoT prompting, Auto-CoT, and Toolformer.

However, they are still far behind humans in their ability to solve complex reasoning problems. To simulate different aspects of human cognition, the OlaGPT model incorporates six cognitive modules (see picture below):

The functions of the cognitive modules are as follows:

1. Intention Enhance (attention) helps the language model to extract the most relevant information.
2. Memory keeps relevant information and brings it back when needed. It stores information in four kinds of libraries: facts, tools, notes, and thinking.
3. Active Learning is inspired by the learning mechanism of humans. When the model fails to solve a problem, the human experts give it a feedback with explanations. The model uses these notes and learns them for similar problems in the future.
4. Reasoning draws conclusions and makes logical connections based on available information. 
5. Controller is used for relevant action selection.
6. Voting (decision making) chooses the best decision for different types of questions (tasks).

The image shows how OlaGPT answers questions by simulating human cognition.

It first analyzes the user’s question intention, and then chooses various cognitive modules, such as thinking templates, relevant resources, tools, facts, and error notes. It further creates and runs different agents that use these modules to find possible answers, and finally selects the best answer based on voting.

The table below shows a comparison between the features used by CoT,  Auto-CoT, Toolformer, and OlaGPT to enhance the LLMs’s reasoning skills.

Experiments and evaluation results

The team did not train OlaGPT from scratch, but used the existing LLMs such as GPT-3 as base models. They designed different cognitive modules and integrated them with the base models.

It was evaluated on two datasets: AQuA, which contains algebraic word problems, and E-KAR, which contains knowledge-intensive analogical reasoning questions in Chinese and English.

They measured the accuracy of the model and compared it with GPT-3.5-turbo, Auto-CoT and SC. They also conduct human evaluation to assess the quality and readability of the chains of thought generated by OlaGPT and the baselines.

The results showed its superior performance over the state-of-the-art benchmarks.

Example of reasoning

Q: What is the next number in the sequence 2, 4, 8, 16, …?

To solve this problem, the model uses the following chain of thought:

1. Recognize that this is a problem of finding a pattern in a sequence of numbers.
2. Use the tool of observation to notice that each number in the sequence is double the previous number.
3. Use the tool of generalization to infer that the pattern is to multiply by 2 each time.
4. Use the tool of application to apply the pattern to the last number in the sequence and get the next number: 16 * 2 = 32.
5. Use the tool of checking to verify that the solution is correct by plugging it into the pattern: 32 / 2 = 16.

Therefore, the answer is 32.

The model’s pipeline for the example above

• Intention Enhance: to focus its attention on the relevant information in the question, such as “the sequence of numbers” and the word ”next”.
• Memory: to recall previous knowledge and facts about sequences and patterns, such as common types of sequences and how to find them. It can store and retrieve relevant information in different libraries: the facts, tools, notes, and thinking libraries. 
• Active Learning: to record the feedback from the user or the expert opinion on my answer, and use it to improve my future performance on similar problems.
• Reasoning: to analyze the sequence and find the pattern, such as observation, generalization, application, and checking.
• Controller: to select the most relevant action for each step of the reasoning process such as to such as observation for step 2 that “each number in the sequence is double the previous number” and multiplication for step 4.
• Voting: to select the best answer based on the confidence and accuracy of the reasoning module, and report it to the user: “the answer is 32”.

Conclusion, future work

OlaGPT is a novel framework that empowers LLMs with human-like problem-solving abilities, simulating certain aspects of human cognition, such as attention, memory, reasoning, learning, and decision making.

It might struggle with novel or complex problems that need creativity, intuition, or common sense, but it is a promising framework that can approximate the human brain model.

For future work, the research team aims to optimize and improve the functions of each sub-module and enhance the model framework. They also plan to use more diverse datasets and baselines.

Learn more:

• Research paper: “OlaGPT: Empowering LLMs With Human-like Problem-Solving Abilities” (on arXiv)
• Project implementation