Вдосконалення агентів LLM шляхом використання їх в якості пошукових агентів

Abstract

This work is devoted to increasing the performance and efficiency of large language models, models that learn the function of producing coherent completion of any given natural language text due to their ability to learn from massive amounts of example data, through test-time scaling. Such models gain many emergent capabilities that make them useful as assistants for professionals, but they also lack some common capabilities of classic systems like search-based agents. We compare common search-based agents to decoder-only large language models to find possible solutions to this problem through additional computation.