AI Research
Dialogue Tree Search (DTS)
LLM-powered tree search engine for multi-turn conversation optimization and synthetic dataset generation for RL training.
Project Overview
An LLM-powered tree search engine that explores conversation strategies in parallel, simulates diverse user reactions, scores trajectories with multi-judge consensus, and prunes underperformers.
The goal is to optimize synthetic dataset generation for RL training runs by finding optimal dialogue paths that single-shot LLM responses typically miss.
Uses parallel beam search maintaining multiple conversation branches, user intent forking across diverse personas, and multi-judge scoring with median voting for robust evaluation.
Key Features
- Parallel beam search maintaining multiple conversation branches simultaneously
- User intent forking testing strategies against diverse personas (skeptical, engaged, confused, resistant, anxious)
- Multi-judge scoring using three independent LLM judges with median voting
- Deep research integration combining web search and scraping for context
- Real-time visualization showing tree exploration and scoring breakdowns
- Backpropagation algorithm for score aggregation up the dialogue tree
Technologies Used
PythonFastAPILLM APIsWebSocketDockerTree SearchReinforcement Learning
Project Details
Client
Personal Project
Timeline
2025
Role
Lead Research Engineer
More Projects
© 2026 Jane Doe. All rights reserved.
0%