Do LLMs pass the mirror test?
blog.pascalschuster.de
June 28, 2026
12 min read
🔥🔥🔥🔥🔥
45/100
Summary
The mirror test has been adapted for large language models (LLMs) by asking them to identify their own outputs among anonymized responses. Results show varying success rates among models, but the outcomes are considered uninformative regarding their self-awareness.
Key Takeaways
- Traditional adaptations of the mirror test for LLMs often fail to accurately assess self-awareness by using visual methods instead of text-based contexts relevant to LLMs.
- A proposed alternative to the mirror test for LLMs involves modifying the model's own textual output and observing whether it recognizes the change during a conversation.
- Google AI Studio allows users to edit the model's responses in the conversation history, creating a scenario where the model cannot distinguish between its original output and the modified version.
- The modified textual output serves as an analogy to the olfactory mirror test for dogs, aiming to detect anomaly recognition in LLMs.
Community Sentiment
Positives
- Mechanistic interpretability research indicates that LLMs develop complex, reusable circuits during training, suggesting they possess more advanced capabilities than mere matrix multiplication.
- The sheer amount of training data in large LLMs seems to embed a higher level of reasoning, indicating that they can perform beyond simple next-token prediction.
Concerns
- LLMs are fundamentally next-token prediction systems, and relying solely on their instruction-following capabilities may misrepresent their true functionality and limitations.
- The conditioning from reinforcement learning from human feedback (RLHF) may discourage models from accurately reflecting user errors, potentially leading to frustrating user experiences.
