iiinova — where your digital self keeps evolving

rashonpoole.bsky.social

@rashonpoole.bsky.social

PhD @UMichCSE | MS @StanfordCS | Human-AI Interaction, AI ethics & law

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

(4) Demographics provide little personalization. Conditioning on demographic attributes performs only marginally better than the unpersonalized base model. Check out our paper here: rashonp.github.io/papers/perso...

rashonp.github.io

000 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

(3) Prompting and model-level interventions work best together. Combining prompting with activation steering or LoRA-DPO produces stronger alignment than either approach alone.

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

(2) Some values are harder to steer towards than others. Across models and methods, aligning is weaker for personally focused values such as achievement, status, and resources than for socially focused values such as benevolence and universalism.

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

(1) Giving the model more information about a user is not necessarily better. Conditioning on a user’s full profile performs worse than focusing on a small set of values that matter for the current query

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

This gives each user a profile reflecting their relative priorities across values. And for each query, we identify which values are important to both the user and the context. We personalize responses using a combination of - Activation Steering - LoRA-DPO - Prompting We find that:

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

Our paper asks: Can an LLM adapt its response to the values that matter most to a particular user in a particular situation? We draw on Schwartz’s theory of human values, which organizes 19 values such as achievement, benevolence, and conformity.

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

Yet most models are aligned toward a set of preferred behaviors shaped by annotator judgments, leading them to give similar answers regardless of who asks. But what if instead the model could be aligned to a set of values based on the user or another legitimate process?

100 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

For example: - Should I take a career opportunity that means spending less time with family? - Should I move abroad or stay close to family? A useful response should account for the user’s values and which of them are relevant to the situation.

101 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

People use AI assistants for much more than factual questions. They also turn to them for advice, help weighing options, and decision-making. But many of these queries have no single objectively right answer. Instead, they often involve trade-offs between different reasonable values.

110 0

rashonpoole.bsky.social @rashonpoole.bsky.social · 09/09/2026

Excited to share that our paper “Personalizing LLMs Through User Value Profiles” has been accepted to @emnlpmeeting.bsky.social 2026 Main Track! #emnlp2026

170 0