The-illusion-of-AGI: Experimental Exploration of Testing the Limits of Large Language Models

The-illusion-of-AGI is an open-source research project aimed at testing the capability boundaries of current state-of-the-art large language models through carefully designed experiments, distinguishing between true intelligence and superficial imitation. The core proposition of the project is: current large language models may create an "AGI illusion"—their performance is convincing, but their actual capabilities are far from AGI. The project uses empirical research to answer key questions such as the nature of LLM understanding, the source of reasoning abilities, domains of strong and weak capabilities, and better testing methods.

The concept of Artificial General Intelligence (AGI) has sparked heated discussions with the excellent performance of LLMs, but is it approaching AGI or just an illusion? The-illusion-of-AGI project explores the true capability boundaries of LLMs with a rigorous experimental spirit. Core propositions include:

• Are LLMs "understanding" or performing complex pattern matching?
• Does their performance stem from reasoning ability or statistical patterns in training data?
• Which tasks do they perform well in, and which expose fundamental limitations?
• How to design tests to distinguish between true intelligence and advanced imitation?

Experimental Design Principles

• Adversarial Testing: Design tasks that expose weaknesses (edge cases, adversarial examples, deep reasoning problems)
• Out-of-Distribution Generalization: Evaluate generalization ability outside training data
• Multi-dimensional Evaluation: Cover dimensions such as understanding, reasoning, creation, common sense, and metacognition
• Human Benchmark Comparison: Quantify the "intelligence gap" between models and humans
• Reproducibility: Open-source experimental code and reproduction guidelines

Key Testing Domains

• Compositional Generalization Testing: Evaluate the ability to reason about concept combinations
• Causal Reasoning Assessment: Distinguish between correlation and causation, counterfactual reasoning
• Physical Common Sense Check: Intuitions about object permanence, gravity, spatial relationships, etc.
• Mathematical and Logical Reasoning: Multi-step reasoning and symbolic manipulation abilities
• Metacognition and Self-Reflection: Confidence assessment, identification of knowledge boundaries
• Long-term Consistency: Consistency of stance and facts in multi-turn dialogues

The project's revealed findings include:

• Superficial Capability Trap: Minor modifications to similar tasks lead to failure, relying on pattern matching rather than understanding
• Impact of Training Data: Capability drops sharply on out-of-distribution tasks, relying on memory rather than reasoning
• Confidence Illusion: High confidence in wrong answers, lack of metacognitive ability
• Limitations in Context Utilization: Short-term context is effective, but long-distance information integration and global consistency are poor
• Creativity vs. Recombination: "Creation" is mostly recombination of elements from training data, not true conceptual innovation

The project prompts a rethinking of the definition of AGI:

• Capability vs. Mechanism: Completing tasks does not mean using human-like mechanisms; AGI should focus on the way it is achieved
• Breadth vs. Depth: Current models have amazing breadth, but there are fundamental limitations in deep understanding and flexible reasoning
• Importance of Robustness: True intelligence needs to remain stable under changes, noise, and adversarial inputs
• Social Embeddedness: Human intelligence relies on social and cultural contexts; is "intelligence" without this considered AGI?

The significance of the project for the AI field:

• Evolution of Benchmark Testing: Promote more rigorous and comprehensive evaluation methods beyond simple accuracy
• Guidance for Research Directions: Identify fundamental limitations of models and point out future challenges to overcome
• Correction of Public Perception: Correct over-optimism or fear about AI capabilities and promote rational discussions
• Safety Considerations: Understanding capability boundaries is crucial for AI safety, clarifying system failure modes

The-illusion-of-AGI adopts an open-source model. The GitHub repository contains experimental code, results, design documents, and contribution guidelines. The community can contribute new test cases, reproduce experiments, propose improvement suggestions, share test results of different models, and accelerate the understanding of LLM capabilities through crowdsourcing.

The project team's future plans:

• Expand test coverage to include more cognitive ability dimensions
• Develop an automated testing framework to support large-scale model evaluation
• Establish a long-term tracking mechanism to monitor changes in model capabilities over time
• Explore testing methods for multi-modal models
• Research capability evaluation in human-AI collaboration scenarios