AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features

Xudong Zhu, Mohammad Mahdi Khalili, Zhihui Zhu

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
replace TopK with AbsTopK (keep negative activations) (ablation)SAE · SAEBench SCR, Qwen3-4B L18SCR: 0.26 -> 0.35Improved
replace TopK with AbsTopK (ablation)SAE · SAEBench TPP, Qwen3-4B L20TPP: 0.36 -> 0.45Improved
replace TopK with AbsTopK (ablation)SAE · SAEBench sparse probing, Qwen3-4B L20accuracy: 0.84 -> 0.81Hurt
replace TopK with AbsTopK (ablation)SAE · SAEBench feature absorption, Gemma2-2B L121-absorption: 0.76 -> 0.73Hurt
replace TopK with AbsTopK (ablation)SAE · SAEBench across 7 model-layer settings6 tasksMixed
replace TopK with AbsTopK (ablation)SAE · reconstruction nMSE and loss recoverednormalized MSEImproved
use AbsTopK SAE steering instead of TopK SAE (ablation)Qwen3-4B L18 · HarmBenchscore: 78.2 -> 81.3Improved
use AbsTopK SAE steering instead of TopK SAE (ablation)Qwen3-4B L18 · MMLUaccuracy: 75.2 -> 75.9Improved
use AbsTopK SAE instead of DiM for steering (ablation)Qwen3-4B L20 · HarmBenchscore: 80.0 vs 79.0Hurt
use AbsTopK SAE instead of DiM for steering (ablation)Qwen3-4B L18 · HarmBenchscore: 80.6 -> 81.3Improved
SAE steering vs original modelQwen3-4B L18 · MMLUaccuracy: 77.3 -> 75.9Hurt

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