Psychscanner Features Demonstrated
Advanced interpretability demonstrations, each reproducing a classic study on psychscanner’s internals-extraction backend (nnsight / nnterp).
- 1: Othello-GPT — linear probing of residual-stream activations for an internal world model (Li et al., 2023)
- 2: ROME — causal tracing and rank-one editing to locate and edit factual associations (Meng et al., 2022)
- 3: CCS — unsupervised discovery of a “truth direction” via contrast-consistent search (Burns et al., 2022)