TLDR Tech / InfoSec
OBLITERATUS
An open-source mechanistic interpretability toolkit that locates and surgically removes refusal behaviors from transformer-based LLMs using SVD decomposition to extract refusal direction vectors from hidden states, then projects them out via norm-preserving biprojection across attention and MLP layers without retraining. The pipeline spans 15 analysis modules covering concept cone geometry, alignment imprint detection (distinguishing DPO vs. RLHF vs. CAI from subspace geometry), cross-model universality indexing, and Ouroboros effect quantification to predict whether guardrails self-repair post-removal. Every opted-in run contributes anonymous benchmark data, including refusal rate, perplexity, and KL divergence, to a crowd-sourced cross-architecture dataset accessible via a community leaderboard on HuggingFace Spaces.
Read the full story