Research map / Architectures and efficiency / RNNs and normalization
Automated Interpretability Metrics Do Not Distinguish Trained and Random Transformers
Thomas Heap, Tim Lawson, Lucy Farnik, Laurence Aitchison
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 trained instead of randomly initialized Pythia weights (ablation) | TopK SAE (R=64, k=32) on Pythia residual stream · Pythia 70M-6.9B, fuzzing AUROC | auto-interpretability AUROC | No clear effect |
| replace consistent token embeddings with i.i.d. Gaussian noise (control) (ablation) | SAE on trained Pythia · Pythia, fuzzing AUROC | auto-interpretability AUROC | Hurt |
| use trained instead of randomized weights (ablation) | SAE on Pythia · Pythia, token-distribution entropy of latents | entropy over token IDs | Improved |
| increase model size (ablation) | SAE on randomized Pythia · Pythia 70m vs 6.9b, fuzzing AUROC | auto-interpretability AUROC gap trained vs random | Mixed |
| train SAEs on 1B tokens instead of 100M (ablation) | SAE on Pythia-70m/410m · Pythia, SAE metrics | evaluation metrics | No clear effect |
| vary expansion factor 16-128 and sparsity 16-32 (ablation) | SAE on Pythia-160m · Pythia-160m | SAE evaluation metrics | No clear effect |
Closest papers by meaning
- Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.92)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.90)
- Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders (NeurIPS 2025 · similarity 0.90)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.90)
- Sparse Autoencoders Do Not Find Canonical Units of Analysis (ICLR 2025 · similarity 0.90)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.90)
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.
