Inside Transformers: Advanced Topics in LLM Interpretability

Reading list

1. Data Attribution — Infini-gram
2. Mechanistic: Localization — Probing — Interpreting GPT: the Logit Lens
3. Mechanistic: Localization — Causality — Locating and Editing Factual Associations in GPT
4. Invited Lecture — Dr. Andreas Waldis presents Aligned Probing: Relating Toxic Behavior and Model Internals
5. Mechanistic: Localization — Neurons — Neurons in Large Language Models: Dead, N-gram, Positional
6. Mechanistic: Localization — Sparse Autoencoders / Transcoders — Sparse Autoencoders Find Highly Interpretable Features in Language Models
7. Application / Evaluation: Unlearning — Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
8. Application / Evaluation: Steering — Steering Llama 2 via Contrastive Activation Addition
9. Application / Evaluation: Safety — Refusal in LLMs is an Affine Function
10. Behavioral — Discovering Latent Knowledge in Language Models Without Supervision
11. Invited Lecture