Research map / Architectures and efficiency / RNNs and normalization
Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects
Hwiyeong Lee, Ingyu Bang, Uiji Hwang, Hyelim Lim, Taeuk Kim
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 |
|---|---|---|---|
| add indirect (module-mediated) effects via full stream transition on key side (ablation) | Logit Lens key variant · GPT-2 Small input score | input score (%): 7.84 -> 39.32 | Improved |
| add indirect effects on value side (ablation) | Logit Lens value variant · GPT-2 Small output score | output score (%): 12.57 -> 15.24 | Improved |
| add indirect effects on value side (ablation) | Logit Lens value variant · Gemma-3-270M output score | output score (%): 6.13 -> 10.63 | Improved |
| use key-side QL for output score (ablation) | QL-Key · GPT-2 Small output score | output score (%): 15.24 -> 1.97 | Hurt |
| use Query Lens instead of Tuned Lens | Tuned Lens · GPT-2 Small input score | input score (%): 8.24 -> 39.32 | Improved |
| use Query Lens instead of Token Change clamping | TC · GPT-2 Small input score | input score (%): 31.47 -> 39.32 | Improved |
| use Query Lens token sets instead of Logit Lens | LL · Qwen-3-1.7B LLM-judged interpretability | interpretability score: 5.06 -> 6.96 | Improved |
| use first-order transition instead of full transition (ablation) | Query Lens · Gemma-3-1B output score layer 22 (near) | output score: 0.156 -> 0.176 | Improved |
| use full transition instead of Logit Lens near output endpoint (ablation) | Query Lens · Gemma-3-1B layer 22 | output score: 0.183 -> 0.156 | Hurt |
| use full transition instead of Logit Lens far from endpoint (ablation) | Query Lens · Gemma-3-1B layer 7 | output score: 0.010 -> 0.046 | Improved |
| increase top-k from 5 to 25 (ablation) | Query Lens · GPT-2 Small input score | input score (%): 30.37 -> 39.32 | Improved |
| QL vs LL at k=5 for output score (ablation) | Query Lens · Qwen-3-1.7B | output score: 6.50 vs 6.65 | No clear effect |
Closest papers by meaning
- Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders (NeurIPS 2025 · similarity 0.86)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.86)
- Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders (ICML 2025 · similarity 0.85)
- ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection (NeurIPS 2026 · similarity 0.85)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.85)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (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.
