1Refusal-Direction Abliteration in Falcon3-1B-Instruct: Runtime Ablation vs. a From-Scratch Constitutional Classifiers++ DefenseSathiyanarayanan Palani·1mo ago·7m readSathiyanarayanan Palani·1mo ago·7m read
1Sparse Autoencoder Emotion Analysis in Gemma 3 1B: Recovering Russell's Circumplex from Disjoint Features vs. Causal Valence-Arousal SteeringSathiyanarayanan Palani·1mo ago·6m readSathiyanarayanan Palani·1mo ago·6m read