Agency of small LLMs in strategic decision-making

Ascertain the extent to which small instruction-tuned large language models can exhibit meaningful agency in strategic decision-making, specifically by selecting and adapting actions in multi-step game-theoretic environments such as the Iterated Prisoner’s Dilemma.

Background

The paper defines agency as strategic decision-making and studies LLMs as agents that choose actions in environments with different outcomes. While large models have been evaluated in various game-theoretic settings, their rationality and strategic reasoning remain debated, and smaller models are widely used in practical deployments.

The authors highlight that whether smaller instruction-tuned models can meaningfully act as agents is unresolved and use Gemma2-2b-it in the Iterated Prisoner’s Dilemma to investigate this question.