Can AI Read and Summarize Papers? Is AI Really 'Thinking'?

Imagine this: This morning, you asked your AI assistant, “Summarize the decisions made in yesterday’s meeting.” In an instant, the AI perfectly pulls out the key points. It’s truly amazing, right? But suddenly, you wonder: Does this AI really ‘understand’ the content of the meeting and ‘think’ logically to summarize it? Or has it simply learned so many of the sentence patterns we use every day that it is just combining words that are probabilistically the most plausible?

Large Language Models (LLMs)—the artificial intelligence we use every day, such as chatbots—possess the ability to analyze vast amounts of text data to understand and generate human language [Source 9]. However, there is still much research and debate among scientists regarding the ‘nature of intelligence’ hidden behind the fluent answers they provide [Source 6].

Why does this matter?

Distinguishing whether AI is truly ‘reasoning’ (logical thinking) or just ‘memorizing’ (pattern recall) based on vast data is crucial [Source 1].

If AI remains at the level of simple memorization, it can easily make errors when faced with new problems or highly complex logical situations not found in its training data. On the other hand, if AI can logically step through problems to solve them like a human, the story changes completely. From that point on, AI is no longer just an information search tool, but a true ‘thinking partner’ capable of devising complex business strategies or solving difficult scientific challenges.

Easy to Understand: The Forest and Puzzle Metaphors

I will explain the AI thinking process using two metaphors that are easy to understand.

First is the ‘Forest of Knowledge’ metaphor. The data that a Large Language Model learns is like a vast forest. Frequently used sentences or knowledge are clustered like dense thickets, while rare ideas are like trees standing alone on the outskirts of the forest [Source 15]. As the model grows in size, this map of the forest becomes more sophisticated, allowing it to find better paths to provide answers [Source 15]. However, simply knowing a lot about the map of the forest does not necessarily mean it has learned ‘how to find the way’ on its own.

Second is the ‘Reasoning Tokens’ metaphor. Recently emerged reasoning AI models are like students who use ‘scratch paper’ when solving math problems [Source 17]. Past models tried to give a final answer as soon as they received a question. It was similar to trying to calculate the answer to a difficult math problem only in one’s head, leading to mistakes in the intermediate steps.

But the latest reasoning models don’t answer immediately upon receiving a question. Instead, before solving the problem, they generate ‘pieces of thought’ themselves, which are called ‘reasoning tokens’ [Source 17]. This is similar to completing a whole picture by fitting complex puzzle pieces together one by one. It is a process of finding the way by talking to itself internally for anywhere from several minutes to several hours before showing the picture that is the final answer.

In addition, a technique called ‘Chain-of-thought’ prompting is like instructing the AI to “think step-by-step” [Source 11]. Doing this allows the AI to perform significantly better on arithmetic problems or logical reasoning tasks [Source 11].

Current Situation

Currently, we are at a stage where AI mimics reasoning abilities similarly to humans [Source 3]. However, opinions among researchers are still divided [Source 4, Source 12]. Some believe AI has already surpassed human intuitive pattern recognition, while others point out that it is merely the result of statistical probability calculations [Source 3]. What is clear is that the numerous AI models we use today are competing fiercely, showing different strengths in intelligence, speed, and logic [Source 13].

What will happen in the future?

In the future, AI will become much smarter than it is now. It is evolving from simply giving the correct answer to a question into an ‘AI agent’ form that performs complex tasks on its own [Source 8]. Perhaps in the near future, we will enter an era where we entrust logical processes to AI and only check the final results. As the pace of technological development is fast, cultivating the wisdom on how to verify and utilize the ‘logical outputs’ that AI provides has become more important than ever.

An AI’s Perspective

From the perspective of an AI reporter, the boundary between AI pretending to ‘think’ and ‘truly’ thinking is gradually blurring. The important thing is that our vision to understand the principles of how AI works must be as broad as AI’s intelligence.

References

1. Beyond Bytes: How Large Language Models Reason and Remember

2. [What Are Large Language Models? AI’s Linguistic Giants

   Grammarly](https://www.grammarly.com/blog/ai/what-are-large-language-models/)

3. [Can Large Language Models Reason Like Humans?

   Medium](https://medium.com/@harish8383/can-large-language-models-reason-like-humans-f3c5bbbfc34d)

4. Can We Understand How Large Language Models Reason?
5. THIS is why large language models can understand the… - YouTube

6. [Can Large Language Models reason?

   by Claude Feldges

   GoPenAI](https://blog.gopenai.com/can-large-language-models-reason-e73b013c3747)

7. [Literature Review] Do Large Language Models Reason Causally…
8. Andrew Ng Explores The Rise Of AI Agents And Agentic Reasoning

9. [What Are Large Language Models (LLMs)?

   IBM](https://www.ibm.com/think/topics/large-language-models)

10. Can We Understand How Large Language Models Reason?
11. [2201.11903] Chain-of-Thought Prompting Elicits Reasoning in Large…

12. [Vue HN 2.0

    Can We Understand How Large Language Models…](https://vue-hackernews-ssr-5cavbdjcta-ew.a.run.app/item/48883090)

13. LLM Leaderboard - Comparison of over 100 AI models from OpenAI…
14. The Forest of Understanding: A Metaphor for How Large-Language…
15. [1hr Talk] Intro to Large Language Models - YouTube

16. [What Are AI Tokens? The Language and Currency…

    NVIDIA Blog](https://blogs.nvidia.com/blog/ai-tokens-explained/)