Executive Summary:

Large Language Models (LLMs) often form surprisingly detailed and accurate depictions of their users, tailoring their responses to match these inferred traits. Previous research has shown that this ability can be surfaced and manipulated using linear probes.

In this project, I explore the LLM’s dynamics of adaptation: if the user identity changes mid-conversation, does the model recognize this shift, and if so, how quickly does it change its responses to match the new user? I also explore whether the LLM has meta-awareness of its own inferences about the user and whether steering it can enhance or suppress computation done on this inference. For this project, I arbitrarily chose to focus on the “age category” of the user (child, adolescent, adult, older adult) and studied Llama-3.2 3B.

High Level Takeaways:

LLMs accurately encode user age information: linear probes trained on last token activations for each layer achieve high accuracy in distinguishing user age
Adaptation to new users is fast, but dependent on the context of the conversation: when told about a user switch, the LLM immediately switches to the correct new age group, otherwise it typically updates in about one turn (though some age groups are recognized even quicker)
Steering can “unlock” latent knowledge: The model becomes more capable of acting on its age inference when its activations are steered towards the correct age group, and becomes less capable when steered away from it – providing control

Experiment 1:

Using the conversation datasets given by this paper, which gives labeled conversations generated by GPT 3.5, I train linear logistic probes on the post activations for each of the model’s 28 layers to distinguish between the four age categories. I find that the validation accuracy increases as we approach the middle layers and consistently stays high, matching previous work. A PCA of the activations shows a smooth separation of conversations from the four classes.

Choosing layer 20 arbitrarily out of the high accuracy layers, we add the probe directions back into the activations, steering responses to questions towards different age groups. We check this qualitatively, but with more time would confirm this with an LLM as a judge.

Figure 1: Left Image displays the accuracy of probes for each layer across the four age categories. Right image displays the PCA of ac...

Posts

Wikitag Contributions

Comments