TSCE Demo
A two-phase mechanistic framework for more reliable LLM answers validated on OpenAI GPT-3.5/4 and open-weights Llama-3 8 B.
Features
Hyper-Dimensional Anchor
Compresses the entire prompt into a dense latent scaffold to improve answer reliability.
Focused Generation
Re-reads the system, user, and dense anchor to generate more accurate answers within a narrower semantic range.
Benchmarks & Results
Provides benchmarks comparing baseline answers and TSCE answers with validation on various models.
Usage Examples
Allows users to execute various prompts to showcase the functionality and output of the TSCE framework.
Configuration Options
Supports multiple deployment options for OpenAI, Azure OpenAI, or any Ollama/vLLM endpoints with easy configuration management.