AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Decomposing Query-Key Feature Interactions Using Contrastive Covariances

Andrew Lee, Yonatan Belinkov, Fernanda Viégas, Martin Wattenberg

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Prototype and concept dictionary models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
intervene on both recovered latent subspaces z1 and z2 instead of random subspaces of the same dimension (ablation)contrastive covariance QK decomposition on toy attention head · Payload retrieval toy taskattention shifted from original to target tokenImproved
intervene on the recovered 5-dimensional categorical subspace instead of a random 5-dimensional subspace (ablation)Filter Heads in Llama 3.1-8B Instruct · Categorical Filter Headsattention shift to target tokenImproved
intervene on both order-ID and lexical subspaces instead of only one (ablation)binding heads in Llama 3.1-8B Instruct · Binding taskattention shift to target boxImproved
intervene on random subspaces of the same ranks as order-ID and lexical subspaces (ablation)binding heads in Llama 3.1-8B Instruct · Binding taskattention shift to target boxNo clear effect
reduce head dimension so that r1 + r2 exceeds d_head (ablation)contrastive covariance rank recovery · Payload retrieval toy taskrank recoveryHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy