LLM Interpretability and Evaluation
Move from a model view to compatible captures, interventions and bounded comparisons.
Read the map first
Start with the model map article to separate architecture, observed activations and causal evidence. Activation mapping then explains token occurrence identity and compatible projections.
Check what a result establishes
Arithmetic across notations requires baseline competence before circuit analysis. Sparse feature recovery separates reconstruction from recovery of known generating features. Glow Worm measures byte prediction rather than chatbot quality.
Related evaluation lessons from security
The signal and protocol studies use different tasks, but make useful comparisons about distribution shift, false alarms and resource budgets. The security catalogue is a separate map of defensive proposals.
Articles in this reading path
- Reading LLM Model Maps: Activations and Causal EvidenceRead a model map without confusing architecture, measured activations and causal evidence, using a local viewer and explicit capture provenance.
- Transformer Activation Maps and Token TrajectoriesTrack individual token occurrences, model revisions and shared projection bases so activation maps remain comparable and their limits stay visible.
- LLM Arithmetic: Digits vs Number WordsA protocol for comparing digit and number word arithmetic circuits, including the failed answer generation gate that blocks causal interpretation.
- Sparse Autoencoders: Reconstruction vs Feature RecoveryA synthetic sparse autoencoder comparison separates reconstruction error from feature recovery using cosine matching, support scores and correlated features.
- Glow Worm: A Small Byte Level Language ModelA small byte level language model tested on WikiText 2, with bits per byte, a unigram baseline and an explicit account of training exposure.
- Synthetic Signal Detection Under Distribution ShiftCompare synthetic signal detection and false alarms under session shift, including a detector whose ranking score hides a failed operating threshold.
- Protocol Resource Limits and Amplification ModellingInspect byte, work, connection and time budgets in local protocol simulations, without treating bounded fixtures as evidence of live vulnerabilities.
- Defensive Security Research ProjectsBrowse defensive research ideas, local implementations and unresolved sources, with practical next checks for incident review and model evaluation.