Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Kolb-Based Experiential Learning for Generalist Agents with Human-Level Kaggle Data Science Performance

Created by
  • Haebom

Author

Antoine Grosnit, Alexandre Maraval, Refinath SN, Zichao Zhao, James Dora, Giuseppe Paolo, Albert Thomas, Jonas Gonzalez, Abhineet Kumar, Khyati Khandelwal, Abdelhakim Benechehab, Hamza Cherkaoui, Youssef Attia El-Hili, Kun Shao, Jianye Hao, Jun Yao, Balazs K egl, Jun Wang

Outline

This paper proposes a computational framework based on Kolb's experiential learning and Vygotsky's zone of proximal development theory, emphasizing that human expertise develops through iterative interaction, reflection, and internal model updating. To overcome the limitations of existing LLM agents' static pretraining and rigid workflows, we design an architecture that separates interaction with the environment (external functions) from internal reflection/abstraction (internal functions). This enables cognitively-based scaffolded learning, enabling open generalization after learning in a structured environment. In a real-world Kaggle data science competition, where the automated data science code generation capability was evaluated across 81 tasks, Agent K achieved a score of 1694, exceeding the median Elo-MMR score of Kaggle Masters (the top 2%), and won multiple medals. This is the first AI system to successfully integrate the cognitive learning theories of Kolb and Vygotsky, representing a significant advance toward general-purpose AI.

Takeaways, Limitations

•
Takeaways:
◦
A new LLM agent learning framework based on Kolb's experiential learning and Vygotsky's zone of proximal development theory is presented.
◦
Agent K demonstrates the potential for developing AI systems capable of solving complex human-level tasks.
◦
Performance verification through experimental results using real data from Kaggle competitions.
◦
Presenting a new direction for the development of general-purpose AI
•
Limitations:
◦
Agent K's performance may be limited to a specific domain (data science).
◦
Further research is needed to determine the generalizability of the proposed framework.
◦
The possibility that it may not be possible to perfectly mimic human cognitive processes
◦
Further discussion is needed on the ethical and social implications.
👍
Made with Slashpage