AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

The Geometry of Updates: Fisher Alignment at Vocabulary Scale

John Sweeney

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language 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
use FisherSketch (joint activation-error geometry) instead of activation-only score on fixed-prefix verbalizer shiftLoRA source selection · Verbalizer shift, Llama-3.1-8BTop-1 accuracy: 20% -> 66.7%Improved
use FisherSketch instead of activation-only marginal on natural domain shiftsLoRA source selection · 100 domains, Llama-3.1-8BTop-1 accuracy: 46.3% -> 45.7%No clear effect
use FisherSketch instead of perplexity similarityLoRA source selection · 100 domains, Llama-3.1-8BTop-1 accuracy: 28% -> 45.7%Improved
add activation geometry to error-only marginal (full FisherSketch) (ablation)FisherSketch · 100 domains, Llama-3.1-8BMax regret: 0.210 -> 0.119Improved
use full error moment instead of diagonal error proxy (ablation)Scov(Γe) · 100 domains, Llama-3.1-8BTop-1 accuracy: 44.0% -> 41.3%No clear effect
use error-only marginal instead of full FisherSketch when activations are constant (ablation)FisherSketch · Verbalizer shift, Llama-3.1-8BTop-1 accuracy: 66.7% -> 72.2%Improved
drop the coupling term and use the separable Kronecker proxy (ablation)FisherSketch · ViT-B/16 same-dataset pairsPearson correlation with exact head Fisher: -0.34Hurt
use CKA instead of exact head Fisher alignment as transfer proxytransfer prediction · ViT-B/16 same-dataset pairsSpearman correlation with exact head Fisher: 0.45 (Pearson 0.43)Hurt
sketch the joint activation-error kernel with random features (m=4096)head Fisher alignment · ViT-B/16, 20 checkpointsspeedup vs exact: 89x speedup at n=2000, m=4096Improved

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