Contrastive neuron attribution locates 0.1% of MLP neurons that distinguish harmful prompts, allowing ablation to reduce refusal rates by over 50% in tuned models while keeping fluency intact.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Targeted Neuron Modulation via Contrastive Pair Search
Contrastive neuron attribution locates 0.1% of MLP neurons that distinguish harmful prompts, allowing ablation to reduce refusal rates by over 50% in tuned models while keeping fluency intact.