DTS: Parallel Beam Search for Dialogue Strategies

Someone built a tool that explores entire conversation trees instead of just generating single responses. It’s called DTS and uses parallel beam search to test dialogue strategies against different user personalities.

How it works:

1. Drop in a goal and opening message
2. It generates N strategies, then forks each against user types (skeptical, cooperative, confused, resistant)
3. Runs full multi-turn conversations down each branch
4. Three LLM judges score trajectories independently, takes the median to filter outliers
5. Prunes weak branches and repeats

The median voting from 3 judges apparently helps a lot with the variance problem when using LLMs as evaluators - outlier scores get filtered automatically.

Grab it here: https://github.com/MVPandey/DTS

Works with OpenAI-compatible endpoints. Fair warning though - it’s token-hungry since it’s exploring multiple conversation paths simultaneously. Pretty useful for researching dialogue approaches or testing how strategies hold up against different user attitudes.