Speaker
Description
Large Language Models have gained widespread
recognition since OpenAI released their revolutionary model,
ChatGPT 3.5. Since then, many new approaches have emerged
to improve the capabilities and accuracy of these models for
different tasks. One such method involves using multi-agent
conversations. This article compares two multi-agent setups
designed to solve the Polish standardized high school exam in
physics. Comparative benchmarks were performed on several
real final exams published by the Polish Central Examination
Board (pl. CKE — Centralna Komisja Egzaminacyjna). The
study employed ChatGPT-4 Turbo and the AutoGen framework.
Benchmarks covered a total of 90 tasks from three Polish Matura
physics exams (editions: 2018, 2019, 2023). The simpler multiagent systems achieved an average score of 76.1%, while the
more complex systems averaged 85.6%.
