A causal audit via neuron-row zeroing shows attribution methods (LRP, IG) faithfully identify dispensable neurons and can install refusal behavior, while rank-stability proxies systematically miss selector failures.
C- ∆Θ: Circuit-restricted weight arithmetic for selective refusal.CoRR, abs/2602.04521, 2026
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
years
2026 2representative citing papers
MANSU achieves the first unlearning method that jointly delivers behavioral forgetting, retain preservation, zero PTQ gap, and structural erasure by circuit-restricted null-space projection plus a per-parameter magnitude floor.
citing papers explorer
-
Faithfulness to Refusal: A Causal Audit of Neuron Selectors
A causal audit via neuron-row zeroing shows attribution methods (LRP, IG) faithfully identify dispensable neurons and can install refusal behavior, while rank-stability proxies systematically miss selector failures.
-
Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution
MANSU achieves the first unlearning method that jointly delivers behavioral forgetting, retain preservation, zero PTQ gap, and structural erasure by circuit-restricted null-space projection plus a per-parameter magnitude floor.