x

LESSWRONG
LW

Ishaan Sinha

Subscribe

Message

9

1

5mo

Do LLMs Change Their Minds About Their Users… and Know It?

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...

Sep 21, 202510

Ishaan Sinha

Subscribe

Message

9

1

5mo

Do LLMs Change Their Minds About Their Users… and Know It?

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...

Sep 21, 202510

Do LLMs Change Their Minds About Their Users… and Know It?

Ishaan Sinha

5mo

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... (read 3977 more words →)

2

10