Can AI plan and reason like humans? A new framework shows promise

To mimic human cognitive processes, the study introduces a thinking method based on a built-in chain of thought (COT). This method enables AI to generate responses by incorporating elements such as chat history, reasoning context, memory, and action planning. The key innovation is allowing models to structure their decision-making process explicitly, making reasoning steps more transparent and improving accuracy.

As artificial intelligence (AI) continues to evolve, researchers are striving to bridge the gap between human cognition and machine reasoning. Despite significant advancements in large language models (LLMs), these models still struggle with multi-turn conversations, contextual reasoning, and efficient action execution.

Addressing these challenges, the research paper "LLM Should Think and Act as a Human" by Haun Leung and ZiNan Wang introduces a novel thinking method based on a built-in chain of thought (COT). Published as a preprint, the study explores how integrating structured reasoning and planning can enhance AI’s ability to process complex requests, reducing errors in multi-turn interactions.

Challenges in multi-turn AI conversations

Current LLMs face difficulties in sustaining logical and accurate responses during multi-turn interactions. One of the main issues is the increasing probability of errors as conversations progress, leading to unreliable outputs. Another limitation is the rigid approach to processing similar user commands. For instance, when asked to help buy a birthday cake, a chatbot may struggle to generate different response workflows based on specific user needs - whether to suggest cake options first or directly confirm the order details. Additionally, existing AI systems often require external tools to execute actions, but the inefficiencies in tool use and limited context memory hinder seamless task execution.

The paper identifies a core reason for these challenges: LLMs lack human-like thinking abilities, such as reasoning, planning, and adaptive execution. To overcome these issues, the authors propose a structured approach where AI processes information similarly to humans - utilizing memory, historical context, and real-time knowledge updates.

A thinking framework based on chain of thought

To mimic human cognitive processes, the study introduces a thinking method based on a built-in chain of thought (COT). This method enables AI to generate responses by incorporating elements such as chat history, reasoning context, memory, and action planning. The key innovation is allowing models to structure their decision-making process explicitly, making reasoning steps more transparent and improving accuracy.

The approach follows a hierarchical process:

1. Contextual Understanding: The model analyzes the conversation history, background context, and relevant knowledge before formulating a response.
2. Reasoning and Planning: AI outlines the necessary steps to fulfill the request, ensuring logical coherence in multi-step tasks.
3. Action Execution: The model interacts with the environment or external tools in a more efficient and structured manner.

To implement this method, the authors trained LLMs using a specialized dataset that emphasizes structured reasoning. Additionally, they introduced a consistency reward model, which refines the AI’s responses through reinforcement learning. This model evaluates how well an AI-generated response aligns with logical expectations, further improving decision-making accuracy over multiple conversation turns.

Enhancing AI efficiency with action calls

One of the study’s notable contributions is the replacement of traditional tool calls with action calls, a more efficient and syntactically elegant approach. Unlike conventional tool calls that occupy significant system context and slow down inference, action calls are dynamically loaded only when needed. This enhances execution efficiency, allowing AI models to seamlessly integrate real-world actions into conversations without overwhelming system memory.

For example, in a scenario where an AI assists with ordering a cake, action calls enable a step-by-step workflow: gathering user preferences, checking availability, and completing the transaction in a structured manner. This method not only optimizes resource usage but also ensures a more human-like interaction flow.

Future implications and conclusion

The research presents promising advancements in AI-driven conversations, making LLMs more adaptable and efficient. By implementing a structured chain of thought and refining action execution, AI can better handle complex, multi-turn tasks. Future developments could focus on expanding these capabilities into broader applications, such as customer service automation, AI-powered assistants, and interactive problem-solving domains.

The study’s findings indicate that AI systems with enhanced reasoning and planning abilities can bridge the gap between machine efficiency and human intuition. As reinforcement learning and structured response modeling continue to evolve, the prospect of AI that truly thinks and acts like a human becomes increasingly feasible.