Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Top Activations: Efficient and Reliable Crowdsourced Evaluation of Automated Interpretability
Tuomas Oikarinen, Ge Yan, Akshay Kulkarni, Tsui-Wei Weng
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 |
|---|---|---|---|
| use model-guided importance sampling instead of uniform sampling (ablation) | correlation estimation · ImageNet simulation | samples to match accuracy: ~15x fewer samples (abstract: ~13x) | Improved |
| use model-guided importance sampling instead of uniform sampling (ablation) | correlation estimation · ImageNet simulation | RCE at same budget: ~65% lower error | Improved |
| add BRAgg (SigLIP prior) rating aggregation instead of majority vote (ablation) | correlation estimation · simulation, 23% noise | raters per input: 2-10x fewer raters | Improved |
| use average aggregation instead of Bayes aggregation (ablation) | correlation estimation · simulation, 23% noise | RCE: fails to reach low error even at 100 raters | Hurt |
| add BRAgg to MG-IS (ablation) | correlation estimation · simulation, 550 evals | evaluations to reach 27.5% RCE: 1760 -> 550 | Improved |
| add MG-IS to BRAgg (ablation) | correlation estimation · simulation | evaluations to reach 27.5% RCE: 14607 -> 550 | Improved |
| combine MG-IS and BRAgg vs uniform + majority vote (ablation) | correlation estimation · simulation | evaluations to reach 27.5% RCE: 22560 -> 550 | Improved |
| add BRAgg to MG-IS (ablation) | correlation estimation · MTurk validation | evaluations to reach 19.8% RCE: 2000 -> 550 | Improved |
| use human study (MG-IS+BRAgg) instead of SigLIP-only evaluation | correlation estimation · all neurons | RCE: 32.10% -> 27.90% | Improved |
| use Linear Explanations instead of LLM-based methods (MAIA, DnD) | neuron explanation · ViT-B-16 layer11 MLP, SigLIP sim | correlation: 0.1497 -> 0.2968 | Improved |
| use LLM-based explanations (DnD, MAIA) instead of simpler CLIP-Dissect | neuron explanation · ResNet-50 layer4, SigLIP sim | correlation: 0.1904 -> 0.1534 (DnD) | Hurt |
| use LLM agent MAIA instead of CLIP-Dissect | neuron explanation · ViT-B-16 layer11 MLP, SigLIP sim | correlation: 0.1849 -> 0.1497 | Hurt |
Closest papers by meaning
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.87)
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.86)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.86)
- A Multimodal Automated Interpretability Agent (ICML 2024 · similarity 0.85)
- Rich Insights from Cheap Signals: Efficient Evaluations via Tensor Factorization (NeurIPS 2026 · similarity 0.85)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 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.
