The Zero Token Transformer (ZTT) introduces a novel approach to maximize the efficiency of existing LLM parameters through an adaptive cycling mechanism, eliminating the need for larger models.

• Enables deeper thinking capabilities by reusing the same parameters in multiple cycles
• Implements a head-tail decoupled parameter cycling method for improved adaptability
• Achieves better performance without increasing model size or computational requirements
• Demonstrates how architectural innovation can overcome resource limitations

This research is particularly valuable for engineering teams working with constrained resources, offering a path to enhance LLM capabilities without the costs associated with larger models.

Zero Token-Driven Deep Thinking in LLMs: Unlocking the Full Potential of Existing Parameters via Cyclic Refinement