AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

The Price of Amortized inference in Sparse Autoencoders

Wenjie Sun, Di Wang, Lijie Hu

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Sparse dictionary representations

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
increase sparsity penalty in standard and TopK SAEsSAE training · Gemma-2-2B layer 12 (SAEBench)NMSE and dead rateHurt
continue training to shrink amortization gapStandard and TopK SAEs · Gemma-2-2B layer 12monosemanticity metricsNo clear effect
add semi-amortized ISTA refinement (ablation)TopK SAE · Pythia-160M layer 8NMSE: 1.499 -> 0.087Improved
use non-amortized ISTA inference (ablation)GatedSAE · Pythia-160M layer 8NMSE: 1.947 -> 0.000Improved
add semi-amortized refinement (ablation)BatchTopK SAE · Pythia-160M layer 8dead latent rate: 0.316 -> 0.000Improved
use non-amortized ISTA inference (ablation)Standard SAE · Pythia-160M layer 8absorption rate: 0.144 -> 0.283Hurt
use non-amortized ISTA inference (ablation)TopK SAE · Pythia-160M layer 8absorption rate: 0.225 -> 0.452Hurt
add semi-amortized refinement (ablation)TopK SAE · TPP Pythia-160M layer 8selectivity: 0.245 -> 0.508Improved
add semi-amortized refinement (ablation)TopK SAE · GIS Pythia-160M layer 8intervention score (KL 1.00): 1.412 -> 1.968Improved
increase ISTA refinement steps from 5 to 50 (ablation)semi-amortized BatchTopK SAE · Pythia-160M layer 8NMSE: 0.477 -> 0.046Improved
increase ISTA refinement steps (ablation)semi-amortized BatchTopK SAE · Pythia-160M layer 8dead rate and feature splitting: dead rate 0.000 throughoutNo clear effect
group latents by angular variance with group encoders (LocA-SAE)amortized SAE · Pythia-160M layer 8dead rate: 0.000 vs 0.316 (BatchTopK full-amortized)Improved

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