Relative Value Encoding in Large Language Models: A Multi-Task, Multi-Model Investigation.

In-context learning enables large language models (LLMs) to perform a variety of tasks, including solving reinforcement learning (RL) problems. Given their potential use as (autonomous) decision-making agents, it is important to understand how these models behave in RL tasks and the extent to which they are susceptible to biases. Motivated by the fact that, in humans, it has been widely documented that the value of a choice outcome depends on how it compares to other local outcomes, the present study focuses on whether similar value encoding biases apply to LLMs. Results from experiments with multiple bandit tasks and models show that LLMs exhibit behavioral signatures of relative value encoding. Adding explicit outcome comparisons to the prompt magnifies the bias, impairing the ability of LLMs to generalize from the outcomes presented in-context to new choice problems, similar to effects observed in humans. Computational cognitive modeling reveals that LLM behavior is well-described by a simple RL algorithm that incorporates relative values at the outcome encoding stage. Lastly, we present preliminary evidence that the observed biases are not limited to fine-tuned LLMs, and that relative value processing is detectable in the final hidden layer activations of a raw, pretrained model. These findings have important implications for the use of LLMs in decision-making applications.