Multi Agent Debate Arena
About the project
Small open-weight LLMs often hallucinate and state wrong answers with confidence, and a single model cannot check itself. We built a locally run debate system where two instances of an open model argue opposing sides of a question, and an independent judge from a different model family picks the stronger answer with a confidence score. Using position-swap checks and length caps to reduce judge bias, we evaluate it against a single-model baseline to show where debate improves reliability, with no data leaving the user’s machine.
Built with
ollamapythonstreamlit
AI tools
Coded withAntigravityLlama 3.3 70B Instruct
Prize Categories Entered
- Submitted

Best Open-Source AI Project