Beyond Token Scale: Chunk-Level Sparse Autoencoders for Reliable Semantic Feature Discovery
cs.CL, cs.AI, cs.LG
Submitted: 2026-09-28
Updated: 2026-09-28
Code: https://github.com/Xu0615/Chunk_Level_SAE
Terminology
Sources
- Layer Normalization
- Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders
- DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
- Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
- Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning
- Applying sparse autoencoders to unlearn knowledge in language models
- The Pile: An 800GB Dataset of Diverse Text for Language Modeling
- Scaling and evaluating sparse autoencoders
- Deep Residual Learning for Image Recognition
- SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models
- Do Sparse Autoencoders Identify Reasoning Features in Language Models?
- Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders
- Representation Learning with Contrastive Predictive Coding
- Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoders
- AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
- Step-Level Sparse Autoencoder for Reasoning Process Interpretation
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering