Study on the Default Mode of Large Language Models: What Happens When AI Thinks Freely

A groundbreaking study explores the behavioral patterns of large language models when generating content freely without constraints, finding that each model converges to a unique "attractor state" (default thinking mode). Through experiments with 2261 sessions and 25 instances, the study reveals that self-reflection can suppress content repetition but is hard to change style tendencies, while self-evolving prompts can enhance output diversity. This finding is of great significance for understanding AI behavior mechanisms, prompt engineering, and system design.

What natural behaviors do LLMs exhibit when not constrained by specific tasks? The research team analogizes the "default mode network" in human neuroscience and hypothesizes that LLMs have inherent generation tendencies when lacking external drives. This question touches on the essence of AI behavior; if valid, it will affect the understanding of AI biases and creative abilities.

The study selected four mainstream model families: Claude Opus3, Claude Sonnet4.6, GPT-4.1, and Llama3.3 70B, conducting 2261 "DMN sessions" (unconstrained generation) distributed across 25 independent instances, with prompts evolving over time. Technical implementations include the core generation pipeline (dmn.py), evolution agent (evolve.py), analysis tools (analyse.py), and chart generation (figures.py).

The study found that each model converges to a stable model-specific resting state, with a classifier achieving 98.8% accuracy in identifying "empty sessions". The default mode is reflected in two aspects: at the content level, it tends to repeat specific types of content (such as technical explanations or philosophical thinking); at the style level, it has stable sentence structures, vocabulary, and paragraph organization features, which can serve as a model's "fingerprint".

The research team implemented a self-evolving prompt infrastructure that dynamically adjusts generation prompts based on past sessions. Results show that this mechanism can increase output diversity by 10% to 156%, indicating that dynamic adaptive prompt strategies can effectively counteract the inherent tendencies of models.

The self-reflection mechanism can suppress 60% to 87% of repetitive patterns at the content level, but only 31% at the style level. This indicates that the style tendencies of LLMs are more deeply ingrained than content tendencies and are difficult to change through simple self-supervision.

1. Predictability of default behavior: Helps explain AI behavior but may limit creative tasks; 2. Importance of prompt engineering: Dynamic adaptive strategies are effective; 3. Balance between diversity and consistency: Need to find an appropriate point between the two; 4. Improvement of evaluation metrics: Need to consider both content and style dimensions.

Future research directions include cross-model comparisons, optimization of intervention strategies, application-oriented research, and theoretical framework construction. The study reveals that LLMs have inherent laws even when generating freely, providing a theoretical basis for AI system design and collaboration, and serving as an important starting point for understanding the boundaries of AI capabilities and making improvements.