Reward model preferences can be explained by generating counterfactual and semifactual answer variations along 15 hand-picked evaluation attributes and measuring which attribute changes flip the model's preference.
Research agenda for sociotechnical approaches to ai safety
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Interpreting Language Reward Models via Contrastive Explanations
Reward model preferences can be explained by generating counterfactual and semifactual answer variations along 15 hand-picked evaluation attributes and measuring which attribute changes flip the model's preference.