Causal interventions in mechanistic interpretability frequently produce divergent representations, including pernicious ones that activate hidden behavioral pathways, and a modified Counterfactual Latent loss can mitigate these while preserving interpretive utility.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Addressing divergent representations from causal interventions on neural networks
Causal interventions in mechanistic interpretability frequently produce divergent representations, including pernicious ones that activate hidden behavioral pathways, and a modified Counterfactual Latent loss can mitigate these while preserving interpretive utility.