Research map / Architectures and efficiency / RNNs and normalization
The Geometry of Updates: Fisher Alignment at Vocabulary Scale
John Sweeney
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use FisherSketch (joint activation-error geometry) instead of activation-only score on fixed-prefix verbalizer shift | LoRA source selection · Verbalizer shift, Llama-3.1-8B | Top-1 accuracy: 20% -> 66.7% | Improved |
| use FisherSketch instead of activation-only marginal on natural domain shifts | LoRA source selection · 100 domains, Llama-3.1-8B | Top-1 accuracy: 46.3% -> 45.7% | No clear effect |
| use FisherSketch instead of perplexity similarity | LoRA source selection · 100 domains, Llama-3.1-8B | Top-1 accuracy: 28% -> 45.7% | Improved |
| add activation geometry to error-only marginal (full FisherSketch) (ablation) | FisherSketch · 100 domains, Llama-3.1-8B | Max regret: 0.210 -> 0.119 | Improved |
| use full error moment instead of diagonal error proxy (ablation) | Scov(Γe) · 100 domains, Llama-3.1-8B | Top-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-3.1-8B | Top-1 accuracy: 66.7% -> 72.2% | Improved |
| drop the coupling term and use the separable Kronecker proxy (ablation) | FisherSketch · ViT-B/16 same-dataset pairs | Pearson correlation with exact head Fisher: -0.34 | Hurt |
| use CKA instead of exact head Fisher alignment as transfer proxy | transfer prediction · ViT-B/16 same-dataset pairs | Spearman 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 checkpoints | speedup vs exact: 89x speedup at n=2000, m=4096 | Improved |
Closest papers by meaning
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
- The Alignment Illusion in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.85)
- Differentially Private Steering for Large Language Model Alignment (ICLR 2025 · similarity 0.85)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (ICML 2025 · similarity 0.85)
- DOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models (ICLR 2025 · similarity 0.85)
- Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing (ICLR 2025 · similarity 0.85)
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.
