Research map / Architectures and efficiency / RNNs and normalization
AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
Xudong Zhu, Mohammad Mahdi Khalili, Zhihui Zhu
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 |
|---|---|---|---|
| replace TopK with AbsTopK (keep negative activations) (ablation) | SAE · SAEBench SCR, Qwen3-4B L18 | SCR: 0.26 -> 0.35 | Improved |
| replace TopK with AbsTopK (ablation) | SAE · SAEBench TPP, Qwen3-4B L20 | TPP: 0.36 -> 0.45 | Improved |
| replace TopK with AbsTopK (ablation) | SAE · SAEBench sparse probing, Qwen3-4B L20 | accuracy: 0.84 -> 0.81 | Hurt |
| replace TopK with AbsTopK (ablation) | SAE · SAEBench feature absorption, Gemma2-2B L12 | 1-absorption: 0.76 -> 0.73 | Hurt |
| replace TopK with AbsTopK (ablation) | SAE · SAEBench across 7 model-layer settings | 6 tasks | Mixed |
| replace TopK with AbsTopK (ablation) | SAE · reconstruction nMSE and loss recovered | normalized MSE | Improved |
| use AbsTopK SAE steering instead of TopK SAE (ablation) | Qwen3-4B L18 · HarmBench | score: 78.2 -> 81.3 | Improved |
| use AbsTopK SAE steering instead of TopK SAE (ablation) | Qwen3-4B L18 · MMLU | accuracy: 75.2 -> 75.9 | Improved |
| use AbsTopK SAE instead of DiM for steering (ablation) | Qwen3-4B L20 · HarmBench | score: 80.0 vs 79.0 | Hurt |
| use AbsTopK SAE instead of DiM for steering (ablation) | Qwen3-4B L18 · HarmBench | score: 80.6 -> 81.3 | Improved |
| SAE steering vs original model | Qwen3-4B L18 · MMLU | accuracy: 77.3 -> 75.9 | Hurt |
Closest papers by meaning
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.92)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.92)
- Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures (NeurIPS 2026 · similarity 0.92)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.92)
- Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.92)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.91)
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.
