Impact of long chain-of-thought priors on multi-turn agentic reasoning

Ascertain how long chain-of-thought predispositions in language models affect performance and tool-use behavior during multi-turn agentic reasoning with external tools.

Background

Long-CoT models often rely on extended internal reasoning and may under-utilize tools, which can hinder agentic RL on reasoning-intensive tasks. The paper later documents this tendency empirically.

A systematic understanding of how Long-CoT priors interact with agentic tool use across tasks and training regimes remains an unresolved issue highlighted by the authors.