Large language models show different decision paths despite similar choices
The Internal Anatomy of Strategic Choice in Large Language Models
Artificial IntelligenceComputer Science and Game Theory
Summary
Sometimes large language models make decisions like humans do, but how they process information inside their 'brains' can be quite different. The researchers looked at how these models play simple games and found that while the end choices looked alike, the steps inside the models varied. Changing how the models were trained affected how information about incentives moved inside but didn’t always change the final decisions. This means similar behaviors in AI can come from very different internal thinking.
What this means in practice
- •For ai system engineers: Diagnose and adjust pathways linking incentives to decisions within language models to refine behavior without changing outcomes.
- •For behavioral economists: Use language models as testbeds to study internal representations of strategic choices and incentives in decision-making scenarios.
Authors
Vinícius Ferraz, Leon Houf, Enrico Ferrea
Abstract
Large language models act as strategic agents and models of human choice, yet choosing like a strategic agent does not mean computing like one. We recorded activations from four open-weight models --- dense and mixture-of-experts, including a matched base--instruct pair --- in one-shot play of 144 strict ordinal $2\times2$ games. We followed a prespecified incentive from prompt, through activations, to choice. Dense models mirrored the unadjusted human decline with game complexity. Incentive and choice were detectable in every model, but models differed in whether incentive reached the choice, aligned with it and, where tested, whether strengthening it shifted preference. The base and instruction-tuned Qwen2.5 models chose almost identically at baseline yet differed in whether incentive reached choice. Fixed decision cues were distinguishable internally but changed choices selectively. Similar behaviour can rest on different computation; post-training can reshape the path from represented incentive to decision while leaving behaviour and decodable information largely intact.