EPA, a contrastive preference loss based on an energy-based model with a unique maximum likelihood estimator, outperforms DPO on offline LLM alignment benchmarks.
Title resolution pending
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
-
Energy-Based Preference Model Offers Better Offline Alignment than the Bradley-Terry Preference Model
EPA, a contrastive preference loss based on an energy-based model with a unique maximum likelihood estimator, outperforms DPO on offline LLM alignment benchmarks.