Optimering av handlingar för stora språkmodeller styrda med spelteori : En undersökning av strategiskt beteende genom utformning av handlingsrymden och spelteoretiska algoritmer

This thesis explores how large language models (LLMs) can be guided to perform better in strategic conversation, specifically in negotiation tasks. While LLMs are proficient at generating coherent text, they lack real strategic reasoning. Recent research has introduced game-theoretical solvers to steer LLMs toward more rational behavior, but the impact of action space design on this process remains understudied. To address this, two types of action spaces, based on emotional tone and explicit offers, were compared to examine how they affect the performance of LLMs when steered with a counterfactual regret minimization (CFR) solver. The action space representations were implemented and paired with a CFR solver in a negotiation environment. The two implementations using different action spaces were evaluated using metrics such as how often they reached agreements, whether the agreements were beneficial, and equilibrium proximity. The results demonstrate that the structure of the action space has a limited effect on strategic outcomes. Both action space representations improved the results, compared to using an uninformed baseline. This suggests that LLMs can understand and utilize both logical actions and emotional tones. The implementation using tones as an action space had a higher percentage of conversations ending with a positive outcome for both parties. However, it also resulted in more conversations with negative outcomes for both parties. This shows that giving the LLM the freedom to make strategic decisions by itself can result in better outcomes, but that there is still a risk of irrational behavior. The work offers insights for designing more effective LLM agents in negotiation and contributes to a broader understanding of how language models can be adapted for complex decision-making tasks.