-->

Research Overview

A joint research team from the Chinese Academy of Sciences (CAS) Institute of Automation and Institute of Neuroscience has confirmed for the first time in the world that multimodal large language models (LLMs) can spontaneously form object concept representation systems very similar to humans. This research result was published online in Nature Machine Intelligence on June 9, 2025, proving that AI has intrinsic understanding of real-world concepts rather than being mere “stochastic parrots.” [Global Times]

Core Research Achievements

Experimental Methodology

The research team conducted the study using a combination of behavioral experiments and neuroimaging analysis:

  • Collected 4.7 million triplet judgment data (targeting LLMs and multimodal LLMs)
  • Derived low-dimensional embeddings capturing similarity structures of 1,854 natural objects
  • 66-dimensional embedding results were stable and predictable, exhibiting semantic clustering similar to human mental representations

Key Findings

Aspect Human AI (LLM)
Concept Formation Visual features + semantic information combined Semantic labels + abstract concept-centered
Consistency Baseline Higher consistency than humans
Dimensional Interpretation Multi-dimensional concept representation Interpretable dimensional structure

Research leader He Huiguang explained, “This study achieved a leap from ‘machine recognition’ to ‘machine understanding,’” describing it as “a core discovery in the ‘mental dimension’ of reaching similar cognitive destinations through different paths.” [Global Times]

Technical Significance and Impact

1. Paradigm Shift in AI Understanding

While existing AI research focused on object recognition accuracy, this study answered the fundamental question of whether models truly “understand” the meaning of objects. It clarified the essential difference between distinguishing cat and dog images and “understanding” that difference.

2. Establishing Cognitive Foundation for Multimodal AI

  • Proved integrated processing capability of language and visual information
  • Elucidated spontaneous formation mechanisms of conceptual representations
  • Confirmed feasibility of implementing human-like semantic clustering

Prospects for the Future Robot World

Short-term Outlook (2025-2030)

1. Rapid Development of Embodied AI Robots

China’s first embodied artificial intelligence robot games scheduled for April 24-26, 2025, in Wuxi will be an important milestone showing how these cognitive breakthroughs apply to physical robots. [Global Times]

2. Emergence of Situationally Aware Robots

  • Contextual understanding: Robots understanding situations and responding appropriately rather than simply executing commands
  • Emotional intelligence: More natural interactions by recognizing human emotional states and intentions
  • Learning capability: Spontaneously forming concepts and adapting in new environments

Medium to Long-term Outlook (2030-2040)

1. New Dimensions of Human-Robot Collaboration

Current: Command → Execution
Future: Intent recognition → Situation understanding → Creative collaboration

2. Industry-specific Innovation Scenarios

Field Current Future Outlook
Healthcare Diagnostic assistance Patient emotion understanding, personalized treatment planning
Education Information delivery Individual learner characteristic recognition, adaptive educational methodology
Manufacturing Repetitive tasks Quality concept understanding, creative problem solving
Service Simple responses Customer need prediction, empathetic service

Long-term Outlook (Post-2040)

1. Emergence of Robot Society

  • Inter-robot collaboration: Different robots cooperating based on common conceptual systems
  • Collective intelligence: Individual robot experiences immediately shared across entire robot network
  • Autonomous social structures: Solving complex social tasks without direct human intervention

2. Ethical Challenges of Human-Robot Coexistence Society

  • Boundaries of consciousness: Philosophical debates about whether robots possess true consciousness
  • Rights and responsibilities: Legal status of robots with human-level cognitive abilities
  • Redefinition of labor: Role division between robots capable of creative thinking and humans

Implications for Korea

1. Need to Bridge Technology Gap

China’s research achievements suggest risks of Korea falling behind in AI fundamental research. Investment expansion in multimodal AI and cognitive science convergence research is urgent.

2. Industry Strategy Review

  • Robot industry: Transition from hardware-centered to cognitive ability-based software
  • AI education: Paradigm shift from simple technology acquisition to cognitive science-based AI understanding
  • R&D: Building convergence research ecosystems between universities, research institutes, and companies

Conclusion

This discovery by Chinese scientists is a historic milestone showing that AI can achieve true conceptual understanding beyond simple pattern matching. This suggests the possibility of robots with cognitive abilities indistinguishable from humans emerging within the next 10-20 years.

Korea must prepare for these changes by expanding fundamental research investment, building convergence research ecosystems, and establishing ethical and legal frameworks for human-robot coexistence society.

Original article: Chinese scientists confirm AI capable of spontaneously forming human-level cognition [Global Times]

참고